by Inference Hub

The Best Open-Source AI Models in November 2026, Ranked by Benchmarks

Kimi K3.1 Pro, GLM-5.3, DeepSeek V4.1 Flash, MiMo V2.6 and more, ranked with independent LMArena and Epoch AI scores, with API prices, licences, the best picks for coding, reasoning and value, and how far they trail the closed frontier.

open-sourcebenchmarksrankingschinese-modelskimideepseekglmqwen

The strongest open-weight model right now is Kimi K3.1 Pro from Moonshot AI. It has the highest Epoch Capabilities Index and the best WebDev coding score of any open model, and costs less than the older Kimi K3, which still edges it in LMArena’s chat ranking. Close behind are Z.ai’s GLM-5.3 and Xiaomi’s MiMo V2.6 Pro.

For value, nothing comes near DeepSeek V4.1 Flash. It’s the second-best open model on Epoch’s index (24th of 258 models, closed ones included) and costs a typical $0.30 per million input tokens and $1.20 per million output tokens.

The gap to the best closed models is real but narrow for chat, wider for coding:

Best open modelBest closed modelGap
Text Arena (chat, human votes)Kimi K3: 1,488 (#22)Gemini 4 Carbon: 1,527 (#1)39 Elo
WebDev Arena (coding)Kimi K3.1 Pro: 1,683Claude Fable 5.2: 1,824141 Elo
Epoch Capabilities IndexKimi K3.1 Pro: 161.2 (#18)Claude Fable 5.2: 173.4 (#1)12 points

The open models are much cheaper, though. Kimi K3.1 Pro’s typical price is $2/$10 per million tokens against $8/$40 for Claude Fable 5.2, and DeepSeek V4.1 Flash costs about a thirtieth of Claude’s price.

All scores are from LMArena and Epoch AI as fetched on 22 November 2026. “Rank” is out of every model each source ranks, closed models included. Prices are the median across the API providers we track; each model page lists them all. Every model below has its weights published on Hugging Face; we checked each one. We’ll update this ranking monthly.

The ranking

Ordered by LMArena’s Text Arena score, which comes from blind head-to-head votes by real users and covers the most models. The Epoch Capabilities Index (ECI) combines dozens of benchmarks into one number; WebDev is LMArena’s coding arena, where people vote on apps the models build.

#ModelMakerText Arena (rank)WebDevECITypical price (in / out, per 1M)
1Kimi K3Moonshot AI1,488 (#22)1,659158.6$3.00 / $15.00
2Kimi K3.1 ProMoonshot AI1,482 (#29)1,683161.2$2.00 / $10.00
3GLM-5.3Z.ai1,482 (#30)1,623157.1$1.40 / $4.40
4MiMo V2.6 ProXiaomi1,481 (#33)1,622–$0.44 / $0.87
5GLM-5.3 FlashZ.ai1,475 (#39)1,610153.8$0.15 / $0.50
6GLM-5.2Z.ai1,475 (#40)1,608153.1$1.40 / $4.40
7DeepSeek V4.1 FlashDeepSeek1,475 (#42)1,658159.4$0.30 / $1.20
8DeepSeek V4 ProDeepSeek1,464 (#65)1,581156.0$1.32 / $3.48
9Kimi K2.6Moonshot AI1,461 (#66)1,509151.6$0.83 / $3.86
10GLM-5Z.ai1,457 (#73)1,435146.5$0.98 / $3.18
11Hy3Tencent1,454 (#82)1,506–$0.14 / $0.55
12Gemma 4 31BGoogle1,452 (#84)1,370144.1$0.14 / $0.40
13Kimi K2.5Moonshot AI1,450 (#87)1,436148.8$0.53 / $2.75
14Step-5StepFun1,449 (#88)1,570–$1.00 / $2.70
15MiMo V2.6 FlashXiaomi1,448 (#92)1,638–$0.14 / $0.28

Text Arena scores within about 10 points of each other are a statistical tie, so places 2–7 are effectively level on chat. The other columns separate them: Kimi K3.1 Pro and DeepSeek V4.1 Flash are clearly ahead on ECI and coding.

Best for coding

LMArena’s WebDev arena, where voters compare working web apps the models build:

ModelWebDevTypical price (in / out)
Kimi K3.1 Pro1,683$2.00 / $10.00
Kimi K31,659$3.00 / $15.00
DeepSeek V4.1 Flash1,658$0.30 / $1.20
Inkling 2 Small1,643$0.30 / $1.00
MiMo V2.6 Flash1,638$0.14 / $0.28
GLM-5.31,623$1.40 / $4.40
MiMo V2.6 Pro1,622$0.44 / $0.87
Qwen3.8 27B1,590$0.45 / $3.00

Two cheap surprises: Inkling 2 Small, Thinking Machines Lab’s Apache 2.0 model from this month, and Xiaomi’s MiMo V2.6 Flash both code nearly as well as Kimi K3 at a tenth of the price or less. This is where the closed models lead most, though: Claude Fable 5.2 and Claude Opus 5.5 score about 1,820.

Best for reasoning and math

Epoch AI’s index plus two of the benchmarks behind it, GPQA Diamond (PhD-level science questions) and OTIS Mock AIME (competition math):

ModelECI (rank)GPQA DiamondOTIS Mock AIME
Kimi K3.1 Pro161.2 (#18)94.2%100%
DeepSeek V4.1 Flash159.4 (#24)93.2%99.4%
Kimi K3158.6 (#26)93.1%97.2%
GLM-5.3157.1 (#34)90.9%91.1%
DeepSeek V4 Pro156.0 (#43)91.7%98.6%
Muse Spark Open155.5 (#44)88.6%98.1%
DeepSeek V4 Flash154.4 (#53)91.0%94.4%
GLM-5.3 Flash153.8 (#54)90.2%93.9%

Meta’s Muse Spark Open is the best open model from a US lab on this measure. LMArena hasn’t ranked it in the Text Arena yet.

Best value

Quality per dollar, using each model’s typical price:

  • DeepSeek V4.1 Flash, $0.30 / $1.20: a top-three open model on ECI and coding. Several providers run launch discounts, with the lowest at $0.06 / $0.24 on OpenScale.
  • GLM-5.3 Flash, $0.15 / $0.50: within 13 Elo of the top open model on chat, with a 1M-token context and an MIT licence.
  • MiMo V2.6 Flash, $0.14 / $0.28: the cheapest model in the top 15, and fifth among open models for coding.
  • DeepSeek V4 Flash, $0.14 / $0.28: older than V4.1 Flash, still 154.4 on ECI.
  • Hy3, $0.14 / $0.55: Tencent’s Apache 2.0 model, 11th on chat.

For a sense of scale: a million output tokens from Claude Fable 5.2 costs $40. The same budget buys about 33 million from DeepSeek V4.1 Flash.

Best to run yourself

The top models are huge (Kimi K3.1 Pro is a 2.8-trillion-parameter mixture of experts), so for a single machine look further down:

  • Qwen3.8 27B (Apache 2.0): 152.6 on ECI and 1,590 on WebDev, the best small model by a distance. About 16GB at 4-bit, so it fits a 32GB computer.
  • Gemma 4 31B and Gemma 4 26B A4B (Apache 2.0): Google’s open models, 12th and 20th on chat among open models.
  • Inkling 2 Small (Apache 2.0): 594B total but only 17B active, so it runs fast on a large-memory workstation.

See which open models you can run locally for memory needs and speeds, and the best mini PCs for local LLMs.

Licences

“Open weights” doesn’t always mean “do anything”. From each model’s Hugging Face page:

LicenceModelsCommercial use
MITDeepSeek V4.1 Flash, V4 Pro and V4 Flash; GLM-5.3 Flash, GLM-5.2; MiMo V2.6 Pro and FlashYes, no conditions beyond keeping the notice
Apache 2.0Qwen3.8 27B, Gemma 4, Hy3, Inkling and Inkling 2 SmallYes, no conditions beyond the notice
Modified MITKimi K3, Kimi K2.6, K2.5Yes; show the model name in your product above 100M monthly users or $20M monthly revenue
CustomKimi K3.1 Pro, GLM-5.3, Step-5, MiniMax M3, Muse Spark Open (access approved by hand), Mistral Medium 3.5Read the licence: most allow commercial use with conditions for very large companies or resellers

For the Chinese labs’ licence terms in more detail, see the licence table in our Chinese open models guide.

What changed since spring

When we last ranked open models in April, the leaders were GLM-5.1, GLM-5, Kimi K2.5 and Qwen3.5 397B, the best of them 36 Elo behind the top closed model. Since then:

  • Moonshot shipped Kimi K3 and, in October, Kimi K3.1 Pro, which took the top spot.
  • Z.ai released GLM-5.2 and GLM-5.3, with four models in the open top 10.
  • DeepSeek moved to V4: the cheap V4.1 Flash beats the bigger V4 Pro on every benchmark here.
  • Xiaomi became a contender with MiMo V2.6.
  • US labs came back to open weights: Google’s Gemma 4, Meta’s Muse Spark Open and Thinking Machines Lab’s Inkling models.

One correction to common listings: Qwen3.7 Max and Qwen3.7 Plus are not open-weight. Alibaba sells them through its API only; its open models this generation are Qwen3.8 27B and the very large Qwen3.8 2.4T A95B, which isn’t benchmarked yet.

To compare any of these models’ prices across providers, open its model page, or browse all models.

Subscribe to our newsletter

News and updates, 1–2 emails per month. No spam.