cd /news/large-language-models/open-weight-models-2-kimi-vs-0-30-de… · home › topics › large-language-models › article
[ARTICLE · art-171597] src=shattered.io ↗ pub= topic=large-language-models verified=true sentiment=· neutral

Open-Weight Models: $2 Kimi vs $0.30 DeepSeek [2026]

Moonshot AI's Kimi K3.1 Pro scores 65.7% on Terminal-Bench 4.0, 23.4 points ahead of DeepSeek V4.1-Flash-1017 at 42.3%, but DeepSeek bills $0.30 per million uncached input tokens at peak versus Kimi's $2.00, according to vendor figures compiled in a Shattered comparison of open-weight models. On the independent Artificial Analysis Intelligence Index, reported by TechDefused, Alibaba's closed Qwen4-Max scored 42.9 points against 31.7 for Z.ai's open-weight GLM-5.3, so the closed model beat the open one on that test. Three of the four leading models ship downloadable weights — Kimi K3.1 Pro, DeepSeek V4.1-Flash and reportedly GLM-5.3 — while Qwen4-Max, released November 16, is available only through Alibaba's QwenCloud API.

recommendation

Weigh token price against benchmark lead before choosing a model: DeepSeek V4.1-Flash charges $0.30 per million uncached input tokens versus Kimi K3.1 Pro's $2.00, so high-volume agent loops pay roughly 6.7x more for Kimi's 23.4-point Terminal-Bench edge.

more in playbook: Controlling LLM cost
by read32 min views6 publishedNov 18, 2026
Open-Weight Models: $2 Kimi vs $0.30 DeepSeek [2026]
Image: Shattered (auto-discovered)

Three of the four leading models in this race ship downloadable weights, and the fourth does not. Kimi K3.1 Pro and DeepSeek V4.1 Flash publish theirs, GLM-5.3 reportedly does too, and Alibaba’s Qwen4-Max (launched November 16) lives only behind a cloud API. On Moonshot’s own tests, Kimi K3.1 Pro scores 65.7% on Terminal-Bench 4.0, 23.4 points ahead of DeepSeek V4.1-Flash-1017 at 42.3%. Yet DeepSeek bills $0.30 per million uncached input tokens at peak, while Kimi bills $2.00.

That gap between score and price defines the open weight models market right now. An independent run on the Artificial Analysis Intelligence Index, reported by TechDefused, put Qwen4-Max at 42.9 points and GLM-5.3 at 31.7, so the closed model beat the open one on that test. Teams weighing the three ways to run open source AI models now have to decide whether weights matter more than a ten-point index lead.

The answer depends on three things: whether you need to host the model yourself, how many tokens your agent loop burns, and which benchmark you trust. Every vendor in this group publishes its own scores, and those scores disagree when two vendors test the same rival. The sections below put the numbers side by side and flag where they clash.

Don't miss new tech stories on Google

Add Shattered once in the Google app and our stories show up in your news suggestions.

Add now

Open Weight Models in November 2026: What Shipped in 60 Days #

A few months ago the open-model conversation centered on Kimi K3 and an older DeepSeek V4 line. Since then, every one of the four labs has shipped something. DeepSeek retired V4 Flash and its vision experiment on September 10 and replaced both with V4.1-Flash, then pushed a 1017 refresh on October 17. Moonshot followed on October 23 with Kimi K3.1 Pro and a faster Highspeed variant. Z.ai positioned GLM-5.3 as its flagship, and Alibaba listed it in a US (Virginia) deployment by November 17. Alibaba’s own Qwen4-Max was announced recently and is available through QwenCloud.

Date (2026) Event Model
September 10 V4.1-Flash released, V4 Flash and V4 Flash Vision Exp retired, new peak/off-peak pricing live DeepSeek
October 17 V4.1-Flash-1017 update across API, web and app, API price unchanged DeepSeek
October 23 Kimi K3.1 Pro and K3.1 Pro Highspeed announced Moonshot AI
October 26 glm-5.3 listed in Alibaba Cloud US (Virginia) scope Z.ai
Early November Kimi K3.1 Pro weights on Hugging Face, Azure and NVIDIA listings dated November 4 Moonshot AI
November 11 Qwen4-Max-Preview, Token Plan access only Alibaba
November 13 DeepSeek opens its full application API suite, GLM-5.3-FlashX-Legal lands in AutoClaw DeepSeek, Z.ai
November 16 Qwen4-Max general release, QwenWork adds 10 models to its picker Alibaba

One successor is announced but not shipped. DeepSeek’s launch page says V4-Pro traffic routes to V4.1-Flash until V4.1-Pro launches, and it gives no date. Treat any V4.1-Pro benchmark you see today as a rumor. Moonshot, for its part, is retiring its K2 line, which the Kimi pricing page says will end by December 2026, so K2.6 and K2.7 Code are history for new projects rather than options.

Release tracking is messy, so expect conflicts. One model tracker lists Kimi K3.1 Pro with a much smaller parameter count than the vendor does. The Hugging Face card and NVIDIA’s model card both describe it as a 3T-class model, and this comparison follows the vendor card. GLM-5.3 shows the same problem with dates: Z.ai’s documentation pages carry different dates depending on the page, while Alibaba Cloud’s changelog shows the model in a US deployment in Virginia by November 17. Without an agreed launch date, treat GLM-5.3 as the current Z.ai flagship and skip any claim about how old it is.

Rankings have not settled either. As the split between open-source LLM leaderboards showed earlier this month, the same model can rank first on one board and fourth on another depending on harness, effort setting and which tasks count. That is why this comparison leans on vendor tables, a cross-check across three sources and one outside index instead of a single leaderboard.

Specs Table: Kimi K3.1 Pro vs DeepSeek V4.1 Flash vs Qwen4-Max vs GLM-5.3 #

The specs below come from each vendor’s own documentation, the Hugging Face model card for Kimi and Alibaba Cloud’s Model Studio page for Qwen4-Max. Where a vendor does not publish a number, the cell says so instead of guessing.

Spec Kimi K3.1 Pro DeepSeek V4.1 Flash (1017) Qwen4-Max GLM-5.3
Maker Moonshot AI DeepSeek Alibaba (Qwen) Z.ai
Newest release Announced Oct 23, 2026 V4.1-Flash Sept 10, 1017 update Oct 17 Nov 16, 2026 (preview Nov 11) Current flagship in Z.ai docs
Weights downloadable Yes, Hugging Face Yes, Hugging Face listing No, API only Reported yes, confirm with Z.ai
Architecture Mixture-of-experts Mixture-of-experts Not detailed in API docs Same base as GLM-5.2, new post-training
Total parameters 2.8T 552B Not published in docs Not stated in sources reviewed
Active parameters 104B 8B (input), 16B (output) Not published in docs Not stated in sources reviewed
Context window 1,048,576 tokens 1M tokens 1,000,000 tokens 1M tokens
Maximum output Up to 1,048,576 (default 131,072) 384K 131,072 (262,144 in thinking mode) 128K
Input types Text, image, video Text and image Text, image, video Text only
Reasoning control Always on, low/high/max Thinking and non-thinking Thinking and non-thinking, low/high/xhigh Always on, low/high/max
API model ID kimi-k3.1-pro deepseek-flash qwen4-max glm-5.3
Faster sibling K3.1 Pro Highspeed, 2 to 3 times faster None listed None listed GLM-5.3-Flash and FlashX
Third-party hosting Azure Foundry, NVIDIA NIM DeepSeek API, web, app Alibaba Model Studio, QwenWork Z.ai API, Alibaba Cloud US scope

Two details deserve attention. First, Kimi is the only model here with a published expert layout: 896 experts, 16 selected per token, two shared experts, and attention built from 69 Kimi Delta Attention layers plus 24 gated DeepSeek Sparse Attention layers, according to the Hugging Face model card. Second, DeepSeek’s asymmetric design uses only 8B active parameters to read input and 16B to write output, which is how a 552B model ends up priced like a small one.

Alibaba’s announcement reportedly describes a 2.4-trillion-parameter backbone with 71 billion parameters active per token, and a third-party Modal listing repeats those figures. Because the Model Studio documentation does not state them, treat them as unconfirmed until Alibaba’s docs do. The same goes for the sharper-sounding claim that Qwen4-Max has 328 billion extra N-gram embedding parameters. It appears in one announcement summary and nowhere else we could verify.

Weights and Licenses: Which Models You Can Actually Download #

An open-weight release lets you download the trained parameters, run them on your own hardware and fine-tune them. It does not guarantee an OSI-style open-source license, and the four contenders differ sharply here. Moonshot ships Kimi K3.1 Pro under its own Kimi K3.1 Pro License, which is custom rather than MIT or Apache 2.0. NVIDIA’s model card for the hosted version says the base model is governed by that license and that use of the NVIDIA-hosted copy falls under NVIDIA’s Open Model Agreement.

Model Weights License status What to check before shipping
Kimi K3.1 Pro Public on Hugging Face, 53,224 downloads in the last month at time of writing Custom Kimi K3.1 Pro License Read the license text for commercial and redistribution terms
DeepSeek V4.1 Flash Hugging Face repo linked from the launch page MIT according to Local Model Watch’s reading of the 1017 repo Confirm the license file in the repo you actually pull
Qwen4-Max Not released Hosted-service terms, fine-tuning and self-deployment marked not supported Alibaba Cloud data-residency terms for your region
GLM-5.3 Reported public by TechDefused Not spelled out on the Z.ai pages we reviewed Verify the download and license on Z.ai’s own repository first

The Qwen row is the plain-spoken one. Alibaba’s Model Studio page marks both model fine-tuning and model deployment as not supported for Qwen4-Max, and TechDefused states that Alibaba keeps the model private and sells it only through Alibaba Cloud. If your compliance team requires on-premises inference, Qwen4-Max is out regardless of its score.

GLM-5.3 is the murkiest. TechDefused says Z.ai publishes the weights, and Z.ai’s documentation calls the model state-of-the-art among open-source models on specific public benchmarks. Neither the pricing page nor the model overview we reviewed shows a download link or a license name. The reasonable reading is that weights exist, but you should confirm before you plan hardware around them. For a wider rundown of terms across more vendors, see this license comparison of six open-weight families.

Why the license matters more than the headline score

A custom license can restrict how you resell inference or embed the model in a competing product. An MIT license almost never does. When a model scores within a few points of its rivals, the license often decides the purchase. A startup that wants to host DeepSeek behind its own API faces a different legal picture than one that hosts Kimi, and a bank that cannot send documents outside its network cannot use Qwen4-Max at all.

Architecture and Context: How Each Model Is Built #

All four models advertise a window of about one million tokens, so context length no longer separates them. What differs is how they spend it. Kimi K3.1 Pro swaps the full-attention layers of its predecessor K3 for gated DeepSeek Sparse Attention layers, which Moonshot says makes long-context modeling more efficient. Its quantization-aware training targets MXFP4 weights and MXFP8 activations, a detail that matters when you size a server.

DeepSeek chose a different route. V4.1-Flash is a 552B mixture-of-experts model with a causal encoder-decoder layout, and DeepSeek says its key-value cache needs a quarter of the high-bandwidth memory and an eighth of the SSD storage of the previous generation. Cache-hit charges often dominate agent bills, so compressing the cache cuts costs directly. The launch page also says the company tested V4.1-Flash against V4-Pro and found the smaller model ahead on performance, cost, speed and total runtime, which is why V4-Pro is being phased out.

GLM-5.3 takes the quietest path. Z.ai says it reuses the GLM-5.2 base model and improves everything through post-training, with environment scaling that now covers tasks resembling several days of work for an experienced engineer. The model always reasons, supports low, high and max effort levels, and rejects requests that try to disable thinking. Qwen4-Max is the multimodal outlier among the closed-weight pair, since Alibaba lists text, image and video input, tool calling, structured output, explicit caching and chain-of-thought summaries.

Reasoning effort is now a price lever

Alibaba’s price table lists output as chain-of-thought plus response, so reasoning tokens cost real money. Kimi and GLM default to max effort, and Qwen4-Max defaults to high. Z.ai’s own figures show why that matters: at High effort GLM-5.3 reaches 31.4% on Z.ai Code Bench using around 50K output tokens, while GLM-5.2 needed 96K tokens to reach 23.4% at Max. Dropping the effort setting for routine work is the cheapest optimization available on all four.

Benchmarks Compared: Three Vendor Tables and One Independent Index #

No neutral lab has published a single head-to-head table of all four models yet. What exists is a set of vendor-authored comparisons plus one outside index, and each has a bias. Moonshot’s Hugging Face card is the widest table and includes Kimi, DeepSeek-V4.1-Flash-1017 and GLM-5.3 side by side. DeepSeek’s own changelog lists its before-and-after scores. Z.ai’s documentation reports GLM-5.3 against its predecessor. The Artificial Analysis Intelligence Index, relayed through OrcaRouter and TechDefused, is the only data point that includes Qwen4-Max.

Benchmark (higher is better unless noted) Kimi K3.1 Pro DeepSeek V4.1-Flash-1017 GLM-5.3
Terminal-Bench 4.0 65.7 42.3 41.8
SWE Marathon v1.1 51.7 40.0 42.5
DeepSWE v1.1 75.5 77.0 66.9
ProgramBench 31.5 29.1 19.0
Kimi Code Bench 3.0 Long-horizon (Moonshot internal) 61.2 51.5 45.5
HLE-Full with tools 70.0 58.8 62.5
Agents’ Last Exam (ALL) 43.2 37.7 26.2
Automation Bench-AA 73.3 66.6 62.2
GPQA-Diamond 94.2 92.2 91.7
HLE-Diamond 39.4 21.6 25.1 (text-only subset)
ExploitBench 74.7 Not reported 54.4
Kimi Hallu Bench Hard, lower is better (Moonshot internal) 8.3 23.3 22.6

Read this table with the author in mind. Moonshot built it, chose the benchmarks and ran Kimi in its own harness, so the right-hand columns carry less weight than the left. Even so, the pattern is consistent. Kimi leads on terminal work and agent suites by 5 to 24 points, DeepSeek edges ahead on DeepSWE v1.1 (77.0 against 75.5), and GLM-5.3 trails on most rows. Anyone building a shortlist from this list of the best open source LLMs should weigh these rows by task, not by the average.

Cross-check: when vendors test the same rival

The sharpest test of a vendor table is whether the other vendors agree with it. For GLM-5.3, three sources report overlapping benchmarks, and the results split in an instructive way.

GLM-5.3 score on Moonshot table DeepSeek table (via Local Model Watch) Z.ai documentation
DeepSWE v1.1 66.9 66.9 66.9
Terminal-Bench 4.0 41.8 Not listed Not reported (Terminal-Bench 3.0: 28.3)
Agents’ Last Exam 26.2 (ALL), 29.4 (CLI) 28.5 28.5
ProgramBench 19.0 19.0 Not reported

DeepSWE v1.1 and ProgramBench match across sources, which suggests those numbers come from a shared technical report. Terminal-Bench 4.0 does not. The vendors’ tables give different GLM-5.3 scores there, which may reflect different harnesses and settings. A model can gain or lose several points on the same benchmark just by changing the agent framework wrapped around it, so single-digit gaps between vendors rarely hold up.

The footnotes in Moonshot’s card spell out how much the setup matters. Kimi runs under the Claude Code harness for Terminal-Bench 4.0, DeepSeek’s own Terminal-Bench figure also uses Claude Code, but its DeepSWE figure uses the Mini-SWE-V2 framework, and several closed-model scores in Moonshot’s table are maximum values pulled from other vendors’ blog posts. Moonshot also notes that proprietary-model testing may involve fallback behavior. None of that makes the numbers wrong. It means each cell answers a slightly different question, so use the table to build a shortlist and your own task set to pick a winner.

The independent index: Qwen4-Max against GLM-5.3

Artificial Analysis Intelligence Index run (OrcaRouter) Qwen4-Max GLM-5.3
Index score 42.9 31.7
Cost to run the full test set $3.27 $5.01
Cost per index point about 7.6 cents about 15.8 cents
Wrong answers on a false-answer trap test 28.4% 38.9%
Tokens read to finish the tests about 10 billion about 10 billion

Qwen4-Max won every ability both models were tested on, according to TechDefused’s summary, and beat GLM-5.3 by 27 points on one hard reasoning test. It cost less despite a slightly higher list price because its cached-token rate is lower. TechDefused also notes that OrcaRouter itself flags the catch: GLM-5.3 can still suit businesses that must run models on their own systems, because Qwen4-Max cannot leave Alibaba Cloud.

Coding and Agent Performance: Where Each Model Wins #

Coding is where the four models separate most clearly, and the story is about generations as much as rivals. DeepSeek’s own changelog shows V4.1-Flash-1017 jumping on every listed agent benchmark against the September release: Terminal-Bench 4.0 from 31.2 to 42.3, DeepSWE v1.1 from 74.2 to 77.0, ProgramBench from 20.3 to 29.1, Agents’ Last Exam from 31.8 to 35.4 and HLE without tools from 36.8 to 42.6. DeepSeek says the update also cut task runtime and output length. For the full score breakdown, see this look at DeepSeek’s Flash-1017 DeepSWE result.

Z.ai reports a similar step for GLM-5.3 against GLM-5.2. Terminal-Bench 3.0 climbs from 4.6 to 28.3, DeepSWE v1.1 from 46.2 to 66.9 and Agents’ Last Exam from 23.8 to 28.5. On Z.ai Code Bench, Z.ai claims a 50% gain over GLM-5.2. Because that benchmark comes from Z.ai itself, treat the result as a vendor claim until outside labs can check it.

Moonshot’s argument is long-horizon work. Kimi K3.1 Pro scores 61.2% on Kimi Code Bench 3.0 Long-horizon, up from K3. That is an internal benchmark. The external numbers are Terminal-Bench 4.0 at 65.7, just behind Claude Opus 5.5 at 66.4 and ahead of GPT-6 Astra at 58.2 in Moonshot’s table. On BenchCAD with tools, which tests 3D reconstruction through CAD code, Kimi scores 92.7% against 96.2% for Opus 5.5 and 95.9% for GPT-6 Astra.

Reasoning, knowledge and hallucination

On pure knowledge tests the four sit close together. Kimi posts 94.2 on GPQA-Diamond against 92.2 for DeepSeek and 91.7 for GLM-5.3, a 2.5-point range. Open weights have closed much of the distance to proprietary models on agent work, though the hardest expert-level reasoning tests are where closed models are most likely to keep an edge.

Factual accuracy is where Moonshot makes its boldest claim. On its internal Kimi Hallu Bench Hard, the claim-level hallucination rate fell from 23.4% for K3 to 8.3% for K3.1 Pro, and DeepSeek and GLM score 23.3 and 22.6 in the same table. Moonshot itself warns that the benchmark is a deliberate stress test and does not represent error rates in daily use. The independent signal points the same direction for Qwen4-Max, which gave 28.4% wrong answers on a false-answer trap test against 38.9% for GLM-5.3.

API Pricing Per Million Tokens: From $0.15 to $20 #

Prices below come from the official Kimi, DeepSeek, Alibaba Cloud and Z.ai pricing pages, checked on November 18, 2026. All figures are US dollars per million tokens. Check the vendor page before you commit, because DeepSeek and Z.ai both adjust prices by time of day.

Model Input (cache miss) Cached input Output Notes
Kimi K3.1 Pro $2.00 $0.20 $10.00 Cache writes $2.00 (5 min) or $4.00 (1 hour), batch at 60% of standard
Kimi K3.1 Pro Highspeed $4.00 $0.20 $20.00 2 to 3 times faster
Kimi K3 (predecessor) $3.00 $0.30 $15.00 Priced 50% above K3.1 Pro
DeepSeek V4.1 Flash, peak $0.30 $0.006 $1.20 Peak 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays
DeepSeek V4.1 Flash, off-peak $0.15 $0.003 $0.60 All other hours, weekends and Chinese holidays
Qwen4-Max (International) $1.50 Listed separately $4.50 Other regions list $1.413 input and $4.24 output
GLM-5.3 $1.40 $0.26 $4.40 Cached-input storage temporarily free
GLM-5.3-Flash $0.15 $0.03 $0.50 Lighter sibling
GLM-5.3-FlashX $0.37 $0.075 $1.25 Targets 200 tokens per second

Kimi K3.1 Pro is cheaper than its predecessor, not pricier. At $2.00 input and $10.00 output it sits a third below K3’s $3.00 and $15.00, so the flagship upgrade comes with a price cut. Access has one condition: the Kimi API unlocks K3.1 Pro after a successful top-up of at least $1, and your cumulative top-ups set your rate-limit tier.

DeepSeek’s schedule has a quirk that favors Western teams. Peak windows run 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, which is 8 p.m. to 11 p.m. and 1 a.m. to 5 a.m. US Eastern Standard Time. A US team working 9 a.m. to 5 p.m. Eastern (14:00 to 22:00 UTC) never touches a peak window, so it pays the off-peak $0.15 input and $0.60 output all day. Z.ai uses a similar scheme for its subscription points, with weekday peak hours of 14:00 to 18:00 Singapore time and half-price points off-peak, including all weekend.

The small tiers deserve a look. GLM-5.3-Flash lists $0.15 input and $0.50 output, which matches DeepSeek’s off-peak input price and undercuts its off-peak output price of $0.60. Z.ai published no benchmark table for the Flash tier in the pages we reviewed, so its quality against DeepSeek V4.1 Flash is an open question. Test it on classification, extraction and drafting before you route real traffic, since those are the jobs a cheap tier usually handles well.

Subscription route: GLM Coding Plan tiers

GLM Coding Plan tier List price per month Displayed discount price 5-hour credits Weekly credits
Lite $18 $12.60 2,000 10,000
Pro $80 $56 12,000 60,000
Max $168 $117.60 28,000 140,000

All three tiers support GLM-5.3 and GLM-5.3-Flash. Z.ai’s pricing page showed FlashX as unavailable on the plan at the time of its last documentation update. Moonshot does not publish an equivalent official subscription list we could verify. A forum user reported a $100 monthly Kimi Max subscription and a $372 annual Pro charge, but those are user reports, and one comes with an allegation that access was never activated, so treat them as anecdotes. For the subscription side of the market, this comparison of GLM Coding Plan, Kimi Code and DeepSeek API pricing goes further.

What a Real Agent Workload Costs on Each Model #

List prices mislead because agents reread the same context hundreds of times. To compare fairly, take one hypothetical workload: 100 million input tokens, of which 20 million are cache misses and the rest are cache hits, plus 10 million output tokens. That mix resembles a coding agent working through a large repository over a few days. The totals below are our arithmetic from the list prices.

Model Cached input Uncached input (20M) Output (10M) Total
DeepSeek V4.1 Flash, off-peak Not listed $3.00 $6.00 Not listed
GLM-5.3-Flash $2.40 $3.00 $5.00 Not listed
DeepSeek V4.1 Flash, peak Not listed $6.00 $12.00 Not listed
Qwen4-Max (International) $6.00 $30.00 $45.00 $81.00
GLM-5.3 $20.80 $28.00 $44.00 $92.80
Kimi K3.1 Pro $16.00 $40.00 $100.00 Not listed
Kimi K3.1 Pro Highspeed $16.00 $80.00 $200.00 $296.00

Three conclusions fall out. DeepSeek at peak costs far less than Kimi K3.1 Pro on this mix, and off-peak, when rates are 50% of peak, the gap widens further. Qwen4-Max comes in below GLM-5.3 on this mix. Finally, Kimi’s cache-write charge can add to its total if every missed token is written at the 5-minute rate, so Kimi’s real total could climb on a cold cache.

Cost per run is not cost per solved task. A cheaper model that needs three attempts can lose to a pricier one that lands the fix once. Kimi’s 23.4-point Terminal-Bench lead over DeepSeek is large enough that, for hard terminal tasks, the 8-to-1 price gap may shrink in practice. For routine work where both succeed, the gap stands in full. Measure retries on your own repository before you pick, and compare more setups in this self-hosted LLM versus API cost breakdown.

Self-Hosting Reality: Memory, Hardware and Inference Engines #

Open weights do not mean cheap to run. Kimi K3.1 Pro is described as a 3T-class open-weight model with about 3 trillion parameters, and a model of that scale is far beyond a gaming PC, a single workstation card or a typical eight-GPU server at full precision.

Moonshot recommends vLLM, SGLang and TokenSpeed for serving Kimi, and NVIDIA’s model card lists Blackwell GB300 hardware with vLLM and NVIDIA Dynamo 1.7.0 as the supported path. Kimi’s weights ship with MXFP4 quantization-aware training, which trims memory against full precision, and the Hugging Face page shows 22 quantized variants and 22 finetunes already. DeepSeek, meanwhile, pitches its own large deployments in terms of 2,000 GPUs plus a storage cluster, and says it will work with the open-source community on inference support, which implies tooling is still maturing. Kimi K3.1 Pro’s move to open weights at 2.8 trillion parameters is a licensing milestone, but a practical one only for teams with cluster-scale hardware.

What a sensible pilot looks like

Start on the vendor APIs, not on hardware. Run 50 to 100 representative tasks through DeepSeek and Kimi for two weeks, log pass rate, retries and tokens per task, and price the result at list rates. If the monthly bill lands well below the cost of a leased GPU node plus the engineer time to keep it patched, stay on the API. If your volume or your data rules push the other way, rent a managed instance of the model you tested before you buy any hardware, since memory needs of this size make a wrong guess expensive.

For most teams the sensible path is a managed host. Kimi K3.1 Pro appears in Microsoft Foundry’s model catalog as a Direct from Azure model with a generally available lifecycle, and NVIDIA lists it on build.nvidia.com with a November 4 release date. Reportedly, DeepInfra also lists it at $1.90 per million input tokens and $9.50 per million output tokens including a 5% promotion, though that is a host price rather than a Moonshot price. Managed hosting gives you data-handling terms you can show an auditor without buying a rack of Blackwell GPUs.

Real-World Examples: Six Deployments, Demos and Warnings #

Benchmarks only say so much, so here are six concrete cases from the past two weeks. Most come from vendor demos, which show capability rather than average behavior, and the list says so where it matters.

  1. QwenWork’s model picker (November 16). Alibaba’s office agent added 10 models to its picker,according to Sulat, citing Leiphone . It includes Qwen4-Max, Kimi-K3.1-Pro, GLM-5.3, GLM-5.3-Flash, DeepSeek-V4-Pro and DeepSeek-V4.1-Flash. QwenWork recommends Qwen4-Max for multi-step reasoning, GLM-5.3 and DeepSeek-V4-Pro for rule and plan checks, and the cheaper Flash models for drafting. The guidance cites no shared benchmark, and enterprise admins can limit models by department, group or member.
  2. Kimi’s walking simulator. In an internal test, Kimi K3.1 Pro built a human walking dynamics simulator on top of the open-source human-atlas model, which has 296 skeletal components and 402 muscle components. Across 167 missions and 420 independent reviews it fixed 36 errors in shoulder-muscle moment-arm calculations and took the simulated human from collapsing within seconds to walking for 11.2 seconds.
  3. Kimi’s game and 3D builds. Moonshot shows K3.1 Pro building a Godot 4.7.2 action-adventure with five explorable regions, an Unreal Engine 5.8 scene with 36 million source triangles and a browser robotic-arm viewer with 2,317 parts and 268 wire routes, verified by 55 passing browser checks. These are curated demos.
  4. DeepSeek’s agent integrations. DeepSeek lists WorkBuddy (including CodeBuddy) and OpenCode as official partners that fully support V4.1-Flash, and it recommends DeepSeek Harness 0.2.1-rc.1 or later for the 1017 model. On November 13 it opened web search, file input and URL reading to developers, with search and fetch priced at $0.75 per 1,000 calls.
  5. GLM-5.3 for cyber and legal work. Z.ai says GLM-5.3 reached its best result to date on the CyberGym vulnerability-discovery benchmark and more than doubled GLM-5.2’s scores on exploitation benchmarks. On November 13 it also put GLM-5.3-FlashX-Legal, a legal variant priced like FlashX, into its AutoClaw product.
  6. Operations reality. Moonshot’s status page logged latency incidents for Kimi Code K2.7 Code Highspeed on November 14 and for K2.8 Preview on November 11 and 17, all marked resolved. A Kimi forum thread from November 16 also reports unexpected credit usage on short prompts for a Max subscription. Neither says anything about K3.1 Pro API quality, but both show why you want a second provider behind your router.

Alibaba’s own cost claims for the closed side are covered in this report on Qwen4-Max and its lower task cost, which is worth reading alongside the TechDefused index run.

Use-Case Recommendations: Which Model for Which Job #

Use case Pick Why
Hardest terminal and long-horizon coding Kimi K3.1 Pro 65.7 on Terminal-Bench 4.0, 61.2 on Kimi Code Bench 3.0 Long-horizon, 75.5 on DeepSWE v1.1
High-volume background agents on a budget DeepSeek V4.1 Flash Off-peak rates are 50% of peak rates, 1M context
Regulated data that cannot leave your network DeepSeek V4.1 Flash or Kimi K3.1 Pro Downloadable weights for self-hosting, though DeepSeek has not officially confirmed an open-weight V4.1 Flash release
Cloud-only API with the best price per index point Qwen4-Max 42.9 index points at about 7.6 cents per point against GLM-5.3’s 15.8
Flat-fee coding subscription GLM Coding Plan Lite from $18 a month, GLM-5.3 and GLM-5.3-Flash on every tier
Interactive chat where latency matters Kimi K3.1 Pro Highspeed or GLM-5.3-FlashX 2 to 3 times faster serving, or a 200 tokens per second target
Video and image understanding through an API Qwen4-Max or Kimi K3.1 Pro Both accept text, image and video input, GLM-5.3 is text-only

Teams rarely need one model. A common pattern is a cheap default with a stronger fallback: route routine edits and summarization to DeepSeek or GLM-5.3-Flash, escalate to Kimi K3.1 Pro when a task fails twice, and keep Qwen4-Max as a second cloud provider for multimodal jobs. The routing sketch in the next section shows one way to wire that up.

Migration Guide: Moving to These Models Without Breaking Calls #

Every model here speaks an OpenAI-compatible protocol, so the swap is mostly a base URL and a model ID. The traps are in the parameters. Each vendor has changed something that breaks old requests.

From Kimi K3 or K2 to K3.1 Pro

  • Change the model to kimi-k3.1-pro and top up at least $1 so the flagship unlocks.
  • Drop temperature ,top_p ,n ,presence_penalty andfrequency_penalty . Moonshot fixes them at 1.0, 0.95, 1, 0 and 0, and requests should omit them.
  • Pass the complete assistant message back in multi-turn and tool-call loops, including reasoning_content andtool_calls .
  • Send images as base64 or ms:// file references because public image URLs are not supported.
  • Plan K2 exits now. New users cannot get K2 models after October 31, and Moonshot plans to retire the series by the end of December 2026.

From DeepSeek V4 Flash to V4.1 Flash

  • Call deepseek-flash and stop hard-coding dated identifiers. DeepSeek says that name always points at the latest version, currently DeepSeek-V4.1-Flash-1017.
  • The old names deepseek-v4-flash anddeepseek-v4-flash-vision-exp still work, but the models behind them are retired and requests are served by V4.1-Flash-1017 at the Flash price.
  • Schedule batch jobs outside the weekday windows of 01:00 to 04:00 and 06:00 to 10:00 UTC to pay half price.

From GLM-5.2 to GLM-5.3

GLM-5.3 always reasons. A request that sets thinking to disabled now fails, so change it before you swap the model ID. Z.ai’s migration notice recommends enabling thinking and setting the effort to low for lightweight calls:

{
  "model": "glm-5.3",
  "thinking": { "type": "enabled" },
  "reasoning_effort": "low"
}

Adding Qwen4-Max as a fallback

Use the model ID qwen4-max. Alibaba maps the effort values minimal to low, medium to high and max to xhigh, defaults to high and uses a thinking budget of 131,072 tokens. Batch inference is listed as not supported in the International scope, and the earlier qwen4-max-preview was limited to Token Plan access. Take the endpoint URL from your Model Studio console for your region.

A minimal router across all four

import os
from openai import OpenAI

PROVIDERS = {
    "cheap": (OpenAI(api_key=os.environ["DEEPSEEK_API_KEY"],
                     base_url="https://api.deepseek.com"), "deepseek-flash"),
    "hard": (OpenAI(api_key=os.environ["MOONSHOT_API_KEY"],
                    base_url="https://api.moonshot.ai/v1"), "kimi-k3.1-pro"),
    "plan": (OpenAI(api_key=os.environ["ZAI_API_KEY"],
                    base_url="https://api.z.ai/api/coding/paas/v4"), "glm-5.3"),
    "cloud": (OpenAI(api_key=os.environ["DASHSCOPE_API_KEY"],
                     base_url=os.environ["QWEN_BASE_URL"]), "qwen4-max"),
}

def ask(tier, messages):
    client, model = PROVIDERS[tier]
    return client.chat.completions.create(model=model, messages=messages)

def solve(messages):
    first = ask("cheap", messages)
    if first.choices[0].finish_reason != "stop":
        return ask("hard", messages)
    return first

Before sending production traffic, run the same 50 real tasks through each tier and record pass rate, retries and tokens per task. Those three numbers turn the 8-to-1 price gap into a cost per solved task, which is the figure that decides whether the cheap tier or the strong tier wins for your workload.

Pros and Cons of Each Model #

Model Pros Cons
Kimi K3.1 Pro Leads vendor tables on terminal and long-horizon coding, takes video input, 2.8T weights are public, price fell a third against K3 Most expensive here at $2.00 input and $10.00 output, custom license, roughly 1.1 TB memory estimate, relies on Moonshot’s own benchmarks
DeepSeek V4.1 Flash Lowest list price, off-peak rates all day for US business hours, 77.0 on DeepSWE v1.1, smaller footprint than Kimi Trails Kimi by 23.4 points on Terminal-Bench 4.0, high internal hallucination score in Moonshot’s table, V4.1-Pro still undated
Qwen4-Max Best price per index point in the one outside test, multimodal input, 1M context No downloadable weights, no fine-tuning, no self-deployment, newest of the four with the thinnest track record
GLM-5.3 Plans from $18 a month, Flash and FlashX tiers, strong cyber and agent gains over GLM-5.2, 1M context Text-only input, lowest scores on most rows of Moonshot’s table, weights and license need confirming, thinking cannot be turned off

Verdict: Which Open Weight Model Wins in 2026 #

No single model wins every column, and the data makes the split clean. Kimi K3.1 Pro is the most capable of the open weight models here. It posts 65.7 on Terminal-Bench 4.0, 61.2 on Moonshot’s long-horizon coding test and 94.2 on GPQA-Diamond, and it comes within 0.7 points of Claude Opus 5.5 on terminal work in Moonshot’s table. DeepSeek V4.1 Flash is the budget pick, with off-peak rates at 50% of peak rates. Qwen4-Max has the strongest outside evidence for quality per dollar but cannot be downloaded. GLM-5.3 is the cheapest way into a flat-fee coding subscription at $18.

Scorecard Kimi K3.1 Pro DeepSeek V4.1 Flash Qwen4-Max GLM-5.3
Terminal-Bench 4.0 (Moonshot table) 65.7 42.3 No vendor comparison published 41.8
Cost of the 100M-token workload Not listed Not listed $81.00 $92.80
Weights you can download Yes Yes No Reported yes
Context window 1,048,576 1M 1,000,000 1M
Image and video input Yes Image Yes No
Lowest entry price to try $1 top-up Pay as you go Alibaba Cloud account $18 a month plan

For most teams the recommendation is a two-tier setup. Run DeepSeek V4.1 Flash as the default, because US daytime traffic lands in its off-peak window and the agent score gap to Kimi is small on repository-level fixes (77.0 against 75.5 on DeepSWE v1.1). Escalate to Kimi K3.1 Pro for terminal-heavy and long-horizon work, where its 23.4-point lead earns the price. Add GLM-5.3 through the Coding Plan if your developers want a fixed monthly bill. Choose Qwen4-Max only when a cloud-only model is acceptable and your jobs include video or image input.

Two cautions apply. First, almost every score above comes from a vendor with a stake in the result, and the Terminal-Bench 4.0 figure for GLM-5.3 moved 3.9 points between two vendors’ tables. Second, closed frontier models still lead on the hardest reasoning, as the HLE-Diamond gap shows, so compare against Kimi K3.1 Pro versus Claude Opus 5.5 if you can use a proprietary model. For wider coverage of this fast-moving field, the AI and machine learning section tracks each release as it lands.

FAQ: Open Weight Models, Kimi, DeepSeek, Qwen4-Max and GLM-5.3 #

Which open weight model is best for coding in November 2026?

On Moonshot’s published table, Kimi K3.1 Pro leads with 65.7 on Terminal-Bench 4.0 against 42.3 for DeepSeek V4.1-Flash-1017 and 41.8 for GLM-5.3. DeepSeek leads on DeepSWE v1.1 at 77.0 against Kimi’s 75.5. Because the table comes from Moonshot, validate the result on your own repository.

Is Qwen4-Max open source or open weight?

No. TechDefused reports that Alibaba keeps Qwen4-Max private and sells it only through Alibaba Cloud, and the Model Studio page marks fine-tuning and self-deployment as not supported.

How much cheaper is DeepSeek V4.1 Flash than Kimi K3.1 Pro?

List input is $0.30 against $2.00 at peak and $0.15 off-peak, and output is $1.20 against $10.00. On a workload with 80% cache hits and 10 million output tokens, that makes DeepSeek far cheaper at peak and cheaper still off-peak.

Can I run Kimi K3.1 Pro on my own hardware?

Technically yes, because the 3T-class weights are public. Practically it needs cluster-scale hardware, so most teams will use a managed host such as Microsoft Azure’s model catalog or NVIDIA NIM.

Are GLM-5.3 weights available?

TechDefused says Z.ai publishes them, but the Z.ai documentation pages we reviewed do not show a download link or license name. Confirm on Z.ai’s own repository before planning a self-hosted deployment.

What changed in DeepSeek-V4.1-Flash-1017?

It is the October 17 update to V4.1-Flash. Terminal-Bench 4.0 rose from 31.2 to 42.3 and DeepSWE v1.1 from 74.2 to 77.0, with shorter runtimes and no price change. Call it through the deepseek-flash model name.

Does DeepSeek’s off-peak pricing apply during US business hours?

Yes. Peak windows are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, so 9 a.m. to 5 p.m. Eastern time falls entirely in off-peak hours at $0.15 input and $0.60 output per million tokens.

When will DeepSeek V4.1-Pro launch?

DeepSeek has not given a date. Its launch page says V4-Pro requests route to V4.1-Flash until V4.1-Pro arrives, so treat any specific date or benchmark as a rumor.

── more in #large-language-models 4 stories · sorted by recency
── more on @moonshot ai 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/open-weight-models-2…] indexed:0 read:32min 2026-11-18 · —