DeepSeek vs Llama vs Qwen

As of July 2026, DeepSeek V4-Pro is the strongest of the three headline open-weight families in this comparison, matching Claude Opus 4.6 on SWE-bench Verified (80.6% vs 80.8%) while costing a fraction as much, whereas Qwen 3.5 wins on multilingual and multimodal breadth and Llama 4 now trails both after Meta stopped shipping open weights. If you widen the field beyond these three, the title of best open-weight model overall is genuinely contested between DeepSeek V4-Pro, Zhipu's GLM-5.2, and Moonshot's Kimi K2.6, all released in the first half of 2026. Llama 4 is the weakest of the three named families and is no longer being advanced: Meta's April 2026 flagship, Muse Spark, is closed-weight and API-only, so Llama 4 Scout and Maverick (April 2025) are Meta's terminal open release.^[1][2][3][4]

This page compares the current flagship of each open-weight family on coding, reasoning, context, license, and cost to self-host, with GLM and Kimi noted as the fastest-rising challengers. It is the open-model counterpart to the frontier flagship comparison. Every number below is vendor-reported unless a source says otherwise, and all figures are stamped for the versions current in July 2026.

Capability matrix: which open model wins each task?

Ratings are grounded in the benchmark table further down; "Best" means it leads or ties the open field on that dimension as of July 2026.

Scroll sideways for more →

Last verified: July 2026.

Standardized spec and license table

Scroll sideways for more →

All five are sparse Mixture of Experts models. Llama 4 also ships a lighter Scout variant (109B total / 17B active, 10M advertised context) and DeepSeek ships a V4-Flash tier (284B / 13B). Qwen3.5-397B-A17B is the largest fully open Qwen; the newer Qwen3-Max and Qwen3.7-Max tiers are proprietary and API-only, so they do not qualify as open weights. Last verified: July 2026.

Benchmark comparison

Canonical benchmark set, vendor-reported unless noted. A blank cell is "n/r" (not reported by the developer), never zero. DeepSeek and Kimi figures use each model's maximum reasoning setting.

Scroll sideways for more →

Sources: DeepSeek V4 technical report and model card ^[1]; Meta "The Llama 4 herd" ^[3]; Qwen3.5 model cards ^[5]; GLM-5.2 model card and release ^[7]; Kimi K2.6 model card ^[9]. Llama 4 Maverick's only coding number is a 43.4 on the October 2024 to February 2025 LiveCodeBench window, which is not comparable to the 2026 v6 harness the others use, and Meta never reported SWE-bench Verified or AIME for Llama 4 because the family has no dedicated reasoning mode.^[3] GLM-5.2 leads the open field on SWE-bench Pro (62.1%, ahead of GPT-5.5 at 58.6%) but Zhipu did not publish a SWE-bench Verified number for 5.2; its predecessor GLM-5.1 scored 77.8 on Verified.^[7][8] Last verified: July 2026.

Which open model is best for coding?

For agentic software engineering the top of the open field is a three-way tie that does not include Llama or Qwen at the very front. DeepSeek V4-Pro resolves 80.6% of SWE-bench Verified issues, statistically level with Claude Opus 4.6 (80.8%) and Gemini 3 Pro (80.6%), and it leads every model, open or closed, on LiveCodeBench v6 at 93.5 with a Codeforces rating of 3206.^[1] Kimi K2.6 is right beside it at 80.2 on SWE-bench Verified and adds an "Agent Swarm" mode that fans a single request out across up to 300 parallel sub-agents.^[9] GLM-5.2 is tuned specifically for long-horizon coding and tops the harder SWE-bench Pro at 62.1%, which several outlets reported as beating GPT-5.5 for roughly a sixth of the cost.^[7]

Qwen3.5-397B-A17B is a clear second tier at 76.4 on SWE-bench Verified, still competitive with mid-2025 frontier closed models.^[5] Llama 4 Maverick is the outlier: Meta never reported SWE-bench Verified for it, and its LiveCodeBench score of 43.4 was measured on an older, easier harness in April 2025.^[3] Because Meta shipped no "Llama 4 Reasoning" model and no extended-thinking mode, Llama 4 is a full generation behind on agentic coding.

Which is best for reasoning and math?

GLM-5.2 posts the single highest competition-math score in the group, 99.2 on AIME 2026, essentially saturating the benchmark.^[7] DeepSeek V4-Pro and Kimi K2.6 both hit 96.4 on AIME 2026, and DeepSeek adds a perfect 120/120 on Putnam 2025 and 90.1 on GPQA Diamond, a PhD-level science exam.^[1][9] GLM-5.2 leads GPQA Diamond at 91.2 and Kimi K2.6 follows at 90.5.^[7][9] Qwen3.5 trails this cluster but is still strong at 91.3 on AIME and 88.4 on GPQA Diamond.^[5]

Llama 4 cannot compete here. With no thinking mode and April 2025 training, Maverick tops out around 88.1 on MATH-500 and 69.8 on GPQA Diamond, and Meta published no AIME result.^[3] On MMLU-Pro, a broad knowledge benchmark, Qwen3.5 (87.8) and DeepSeek V4-Pro (87.5) lead the group, with Maverick at 80.5; GLM-5.2 and Kimi K2.6 did not publish a text MMLU-Pro figure (Kimi reports 79.4 on the multimodal MMMU-Pro, which is a different test).^[1][3][5][9]

Which has the biggest and most usable context window?

On paper Llama 4 Scout wins with a 10-million-token advertised window, but independent testing found both Scout and Maverick degrade sharply past a few hundred thousand tokens because the chunked-attention scheme creates blind spots across chunk boundaries.^[3] The most usable long context belongs to DeepSeek V4-Pro and GLM-5.2, each rated at a full 1,000,000 tokens. DeepSeek's hybrid Compressed Sparse Attention cuts inference cost at 1M tokens to about 27% of its previous generation, which Hugging Face framed as "a million-token context that agents can actually use," and it scores 83.5 MMR on the MRCR 1M needle-in-a-haystack test.^[1] GLM-5.2 also serves a 1M window with up to 131,072 output tokens.^[7]

Qwen3.5-397B-A17B ships a native 262,144-token (256K) window using a linear-attention Gated DeltaNet stack, extendable toward roughly 1M with YaRN scaling; the hosted Qwen3.5-Plus endpoint exposes 1M by default.^[5] Kimi K2.6 holds a 256K window.^[9] For workloads that genuinely reason over 500K-plus tokens, DeepSeek V4-Pro is the safest open choice.

Which license is actually the most open?

This is where Llama 4 falls furthest behind. DeepSeek V4 (MIT), GLM-5.2 (MIT), and Qwen3.5 (Apache-2.0) all carry standard, OSI-recognized permissive licenses with no user-count ceiling, no acceptable-use policy, and no naming requirement.^[1][5][7] Kimi K2.6 uses a "modified MIT" that is close behind.^[9] Llama 4, by contrast, ships under the Llama 4 Community License, a source-available license that is not OSI-approved: it requires a separate commercial license from Meta once a product exceeds 700 million monthly active users, attaches an acceptable-use policy, and mandates "Built with Llama" attribution.^[3] For most developers those terms are harmless, but for large commercial deployments the Chinese open-weight models are legally simpler to build on.

What does it cost to run each model?

Because all five publish downloadable weights, the true cost of "self-hosting" is GPU rental, not per-token fees: renting an 8-GPU H100 or H200 node runs on the order of $15 to $30 per hour, and a 1T-parameter MoE like Kimi K2.6 or DeepSeek V4-Pro needs a full node, while Qwen3.5-397B or a quantized GLM-5.2 fits in less. The table below instead lists the lowest widely available third-party hosted price for each open model, which is the practical market cost if you do not own hardware.

Scroll sideways for more →

Prices in USD per 1,000,000 tokens, as of July 2026. Llama 4 Maverick and DeepSeek V4-Flash are the cheapest capable options; Qwen3.5-Plus is the cheapest of the current flagships. Even the priciest here, GLM-5.2 and Kimi K2.6, undercut comparable closed models such as GPT-5.5 ($5.00 / $30.00) and Claude Opus 4.6 ($15.00 / $75.00) by roughly six to twenty times on output.^[2][11]

Where does Meta stand on open weights in 2026?

Meta's open-weight strategy effectively ended between April 2025 and April 2026. Llama 4 Behemoth, the ~2T-parameter teacher model announced in April 2025, was never publicly released after the team hit MoE-routing and chunked-attention instability at 2T scale, and Meta never issued a formal cancellation.^[3] On April 8, 2026, the newly formed Meta Superintelligence Labs shipped Muse Spark, a proprietary, API-only multimodal reasoning model described as Meta's first frontier model, released "without open weights, a sharp break from the open LLaMA lineage."^[4] The practical consequence for anyone choosing an open stack: Llama 4 Scout and Maverick are Meta's last open-weight models for the foreseeable future, and the frontier of open weights has moved decisively to Chinese labs (DeepSeek, Alibaba, Zhipu, Moonshot).

GLM-5.2 and Kimi K2.6: the challengers

Two 2026 releases now rival or beat DeepSeek and Qwen. GLM-5.2, from Zhipu AI under the Z.ai brand, is a ~753B-parameter MoE (about 40B active) released June 13, 2026 under MIT with a 1M-token context, purpose-built for autonomous coding agents; it tops the open field on SWE-bench Pro (62.1%) and AIME 2026 (99.2).^[7] It was trained entirely on Huawei Ascend hardware with no Nvidia chips, a point widely covered as evidence that export controls have not stopped frontier-tier Chinese models.^[8] Kimi K2.6, from Moonshot AI, is a ~1T-parameter MoE (about 32B active) released April 20, 2026; at launch it ranked #1 among open-weight models and #4 overall on the Artificial Analysis Intelligence Index with a score of 54, trailing only Anthropic, Google, and OpenAI, and it is natively multimodal with an Agent Swarm mode.^[9][10] Both are more current than Llama 4 Maverick and, on their strongest axes, ahead of Qwen3.5.

Which should you choose?

• Best overall open model for coding and reasoning, cheaply: DeepSeek V4-Pro. It ties Claude Opus 4.6 on SWE-bench Verified, leads on LiveCodeBench, serves a genuinely usable 1M context, and ships under MIT. Use DeepSeek V4-Flash when cost matters more than the last few points.^[1][2]
• Best for agentic, long-horizon coding: GLM-5.2, which leads SWE-bench Pro and runs autonomous sessions for hours, or Kimi K2.6 for its Agent Swarm and top-tier all-round score.^[7][9]
• Best for multilingual, multimodal, or on-fewer-GPUs deployment: Qwen3.5-397B-A17B. It covers 201 languages, accepts image and video natively, activates only 17B parameters, and is Apache-2.0.^[5]
• Choose Llama 4 only if you are already invested in the Llama ecosystem or need its specific tooling; on benchmarks, license terms, and recency it now trails all four Chinese options, and Meta has moved its frontier work to the closed Muse Spark.^[3][4]

The short version: for the named three-way contest, DeepSeek V4-Pro beats Qwen3.5, which beats Llama 4 Maverick. For the true "best open model" question, DeepSeek V4-Pro, GLM-5.2, and Kimi K2.6 are the ones to shortlist.

References

1. DeepSeek, "DeepSeek-V4-Pro technical report and model card," Hugging Face (deepseek-ai/DeepSeek-V4-Pro), April 2026. ↩
2. DeepSeek, "API pricing," platform.deepseek.com, 2026 (V4-Pro standard $1.74 / $3.48 per 1M; V4-Flash $0.14 / $0.28). ↩
3. Meta AI, "The Llama 4 herd: the beginning of a new era of natively multimodal AI innovation," ai.meta.com/blog/llama-4-multimodal-intelligence, April 5, 2025 (Scout/Maverick specs, benchmarks, license). ↩
4. Meta Superintelligence Labs, "Introducing Muse Spark," April 8, 2026 (Meta's first closed-weight frontier model). ↩
5. Qwen Team, Alibaba, "Qwen3.5: Towards Native Multimodal Agents," qwen.ai, plus Qwen3.5-397B-A17B Hugging Face model card, February 16, 2026. ↩
6. Alibaba Cloud, "Qwen3-Max and Qwen3.7-Max," Model Studio (proprietary, API-only flagship tiers), 2025 to 2026.
7. Z.ai (Zhipu AI), "GLM-5.2 model card and release," Hugging Face, June 13, 2026; VentureBeat, "Z.ai's open-weights GLM-5.2 beats GPT-5.5 on multiple long-horizon coding benchmarks for 1/6th the cost," June 2026. ↩
8. Z.ai (Zhipu AI), "GLM-5.1 model card," Hugging Face, April 7, 2026 (SWE-bench Verified 77.8; Huawei Ascend training). ↩
9. Moonshot AI, "Kimi-K2.6 model card," Hugging Face (moonshotai/Kimi-K2.6) and kimi.com/blog/kimi-k2-6, April 20, 2026. ↩
10. Artificial Analysis, "Intelligence Index" leaderboard, April 2026 (Kimi K2.6 = 54, #1 open-weight, #4 overall). ↩
11. OpenRouter model pricing pages for Llama 4 Maverick, Qwen3.5-Plus, GLM-5.2, and Kimi K2.6, accessed July 2026. ↩
12. Moonshot AI, "Kimi API pricing," platform.moonshot.ai, 2026 (K2.6 $0.95 input / $4.00 output; cache-hit input $0.16 per 1M). ↩