What are the best open-weight models for coding agents in October 2026?
GLM 5.3 is the strongest open-weight model for coding agents in October 2026, with the highest open-weight result on the official Terminal-Bench 4.0 leaderboard. GLM 5.3 Flash and DeepSeek V4.1 Flash handle routine steps for a fraction of the price, and Kimi K3 and MiMo V2.6 Pro are the other large options. None of them matches the frontier models Claude Fable 5.1 and GPT-6 Astra on long agentic tasks. Standard Compute serves all of them on one API key, and its router sends each step to the cheapest model that can do it. Llama and gpt-oss are easier to self-host but well behind for agent work.
The open-weight models we serve, by job
Every model below has its weights published on Hugging Face and is served on Standard Compute. The order is our recommendation for agent work: three models for hard steps, three for routine steps, then three for narrower uses, with our pick first in each group. The reasons are in the next section. Two closed frontier models are listed underneath for scale.
| Model | Use it for | Weights license | Context | Terminal-Bench 4.0 | Price per 1M in / out |
|---|---|---|---|---|---|
| GLM 5.3 | Hard steps | GLM 5.3 License | 1M | 41.8% | $1.40 / $4.40 |
| Kimi K3 | Hard steps | Kimi K3 License | 1M | Not listed | $3 / $15 |
| MiMo V2.6 Pro | Hard steps | MIT | 1M | Not listed | $0.435 / $0.87 |
| GLM 5.3 Flash | Routine steps | MIT | 1M | 35.8% | $0.15 / $0.50 |
| DeepSeek V4.1 Flash | Routine steps | MIT | 1M | Not listed | $0.30 / $1.20 |
| MiMo V2.6 Flash | Routine steps | MIT | 1M | Not listed | $0.14 / $0.28 |
| Qwen 3.8 27B | Narrower uses | Apache 2.0 | 262K (up to 1M) | Not listed | $0.50 / $3 |
| MiniMax M3 | Narrower uses | MiniMax Community License | 1M | Not listed | $0.30 / $1.20 |
| DeepSeek V4 Pro 0813 | Narrower uses | MIT | 1M | Not listed | $1.32 / $3.96 |
| Claude Fable 5.1 (closed) | Frontier, for scale | Proprietary | 1M | 57.9% | $10 / $50 |
| GPT-6 Astra (closed) | Frontier, for scale | Proprietary | 1.05M | 58.2% | $10 / $50 |
Terminal-Bench 4.0: the official leaderboard at tbench.ai, read October 15, 2026 (GLM models and Claude Fable 5.1 run in Claude Code, GPT-6 Astra in Codex); "Not listed" means the model has no entry there yet. Licenses and context: each model's Hugging Face card and license file (linked), and Z.ai's docs for GLM 5.3 and GLM 5.3 Flash. Prices: each maker's own pricing page, read Oct 14, 2026 (all 32 served models); DeepSeek's are peak-hour rates, halved off-peak, and MiniMax doubles above 512K input.
Which one for which job
- GLM 5.3,
z-ai/glm-5.3: Our open-weight pick for hard steps such as multi-file changes and long tool loops. It has the highest open-weight result on the official Terminal-Bench 4.0 leaderboard (41.8%, run in Claude Code). - Kimi K3,
moonshotai/kimi-k3: Moonshot's 2.8-trillion-parameter flagship, with image and video input. At $3 / $15 per million tokens it costs like a closed mid-tier model, so keep it for planning turns, not every step. - MiMo V2.6 Pro,
xiaomi/mimo-v2.6-pro: Xiaomi's 1.02-trillion-parameter flagship at a low per-token price. Released September 21, it is not on the official Terminal-Bench leaderboard yet and has little agent track record, so test it before relying on it. - GLM 5.3 Flash,
z-ai/glm-5.3-flash: Our pick for routine steps, subagents and screenshots (it takes images): 35.8% on the official Terminal-Bench 4.0 leaderboard, the second open-weight entry, at $0.15 / $0.50 per million tokens. - DeepSeek V4.1 Flash,
deepseek/deepseek-v4.1-flash: High-volume edits, tests and lookups. #1 by tokens on OpenCode's public usage board on October 13. - MiMo V2.6 Flash,
xiaomi/mimo-v2.6-flash: The lowest per-token price here, for routine steps. New since September 21, like MiMo V2.6 Pro. - Qwen 3.8 27B,
qwen/qwen3.8-27b: The only model here small enough to self-host on one workstation. Suited to completions and light edits rather than long agent runs. - MiniMax M3,
minimax/minimax-m3: Fast and cheap with a 1M context, but its license is the most restrictive here, and it is not on the official Terminal-Bench leaderboard. - DeepSeek V4 Pro 0813,
deepseek/deepseek-v4-pro-0813: Mostly superseded for agents by the newer V4.1 Flash, which costs a fraction as much per request in our measurements. Still DeepSeek's largest open model (1.6 trillion parameters).
How close are they to Claude Fable 5.1 and GPT-6 Astra?
Not level yet. On the official Terminal-Bench 4.0 leaderboard, the best open-weight result is GLM 5.3 at 41.8%, against 58.2% for GPT-6 Astra and 57.9% for Claude Fable 5.1. Most open-weight models have no entry there at all.
The price gap is far larger than the quality gap. On the coding-agent requests we measured, an average request costs $0.0067 on DeepSeek V4.1 Flash and $0.057 on GLM 5.3, against $0.24 on Claude Fable 5.1 and $0.41 on GPT-6 Astra, all at the makers' list prices (how we measure).
In practice that splits the work: open models handle the routine majority of agent steps (edits, test fixes, file reads, summaries), and a frontier model earns its price on the steps where a wrong plan wastes the next fifty calls, such as cross-service refactors and debugging that has already failed twice.
What the licenses allow
- MIT: DeepSeek V4.1 Flash, DeepSeek V4 Pro 0813, GLM 5.3 Flash and the MiMo V2.6 checkpoints. Use, modify and serve them commercially, keeping the copyright notice.
- Apache 2.0: Qwen 3.8 27B. The same freedom, plus an explicit patent grant. Alibaba's hosted Qwen 3.7 Plus, Qwen 3.7 Max and Qwen 3.8 Max are API-only, not open-weight.
- GLM 5.3 License: MIT-style, but a company running a model-as-a-service business with more than $10 billion in revenue over 12 months must pass Z.ai's security review before commercial use.
- Kimi K3 License: MIT-style, but a model-as-a-service business above $20 million in revenue over 12 months needs a separate agreement with Moonshot, and a product above 100 million monthly users or $20 million monthly revenue must show "Kimi K3" in its interface. Internal use is exempt.
- MiniMax Community License: the most restrictive here. Commercial use requires "Built with MiniMax M3" on your site or docs and a notice to MiniMax, and written authorization above $20 million in yearly revenue.
- These terms bind whoever runs the weights. Calling a hosted API falls under that API's terms instead. This is a summary, not legal advice.
Self-host or call an API?
Most of these models are far too large for one machine: Kimi K3 has 2.8 trillion parameters, MiMo V2.6 Pro 1.02 trillion, DeepSeek V4 Pro 1.6 trillion, and even the Flash models are 300B to 550B, so they need a multi-GPU server. Qwen 3.8 27B is the exception, and a local model also has to be fast enough for an agent loop that makes dozens of calls per task.
Hosted endpoints decide price, speed, precision and where your code is processed. Several hosts serve open-weight models at reduced precision (FP8 or FP4); which ones we serve that way.
Llama, gpt-oss and Mistral Large 4
- Llama 4 Maverick (Meta, Llama 4 Community License) and gpt-oss-120b (OpenAI, Apache 2.0) are the easiest well-known models to run yourself; OpenAI says gpt-oss-120b fits on a single 80 GB GPU. For agents they are far behind: neither is on the official Terminal-Bench 4.0 leaderboard, and Artificial Analysis records 0% for both in its own Terminal-Bench 4.0 run (Source: Artificial Analysis (artificialanalysis.ai)). We don't serve either.
- Mistral Large 4 (public preview since October 6) is described by Mistral as open-weight, but its weights and license are not on Hugging Face yet. It is on Standard Compute as
mistralai/mistral-large-4-0; it joins the table when the weights ship.
Run them in your agent through one API
Standard Compute is one OpenAI- and Anthropic-compatible API with every model above. No Z.ai, Moonshot, MiniMax, Xiaomi or Alibaba account is needed.
- Cost first: select up to 5 models in the dashboard, for example GLM 5.3 Flash and DeepSeek V4.1 Flash for routine work, GLM 5.3 and Claude Fable 5.1 for hard steps, and send the model
standardcompute. The router sends each step to the cheapest of them that can do it. - For every model we vet providers for price, uptime and speed, with automatic fallback when one fails. A request with an exact model ID, such as
z-ai/glm-5.3, always stays on that model. - Opt-in Compliance rules restrict a key, team or company to provider endpoints in Europe or the US. The table below shows which open-weight models each rule keeps; a model without an endpoint in the region fails before any prompt is sent.
- We don't train on your prompts or store them.
Which open-weight models each region rule keeps
| Compliance rule | Open-weight models it keeps |
|---|---|
| Europe | GLM 5.3 and GLM 5.3 Flash |
| United States | GLM 5.3, Kimi K3, GLM 5.3 Flash, DeepSeek V4.1 Flash, Qwen 3.8 27B, MiniMax M3 and DeepSeek V4 Pro 0813 |
| Neither yet | MiMo V2.6 Pro and MiMo V2.6 Flash: requests under either rule fail |
From the endpoint list the gateway enforces, reviewed October 13, 2026. Without a rule, every model routes across all its vetted providers.
Setup in OpenCode, Claude Code and Codex
- OpenCode: run /connect, choose Standard Compute, paste your key and pick a model such as
z-ai/glm-5.3-flash. - Claude Code: set ANTHROPIC_BASE_URL to https://api.stdcmpt.com, ANTHROPIC_AUTH_TOKEN to your key, and ANTHROPIC_MODEL plus each ANTHROPIC_DEFAULT_*_MODEL slot to a model ID. GLM 5.3 in Claude Code and Kimi K3 in Claude Code have tested slot mappings.
- Codex:
standardcompute launch codex -- --model deepseek/deepseek-v4.1-flash. - Cline, Kilo Code, Hermes and other OpenAI-compatible agents: base URL https://api.stdcmpt.com/v1 with the same key and model IDs.
The catch
Open weights don't make a model cheap, private or good at agent work by themselves. The host you call decides the price, the precision and where your code goes, and the best open-weight model still loses to Claude Fable 5.1 and GPT-6 Astra on long, multi-step tasks. Use open models for the volume and keep a frontier model for the steps that decide the outcome.
Where Standard Compute fits
One key for every open-weight model above plus Claude, GPT and Gemini, with a router that picks the cheapest capable model per step. Plans start at $19 a month, as a fixed compute budget.
FAQ
Which open-weight model is best for Claude Code?
GLM 5.3 for the main model and GLM 5.3 Flash for background calls and subagents is the strongest open-weight pairing. For the lowest cost, DeepSeek V4.1 Flash in every slot works. Set ANTHROPIC_BASE_URL to https://api.stdcmpt.com and the model slots to z-ai/glm-5.3 and z-ai/glm-5.3-flash.
Is GLM 5.3 better than DeepSeek V4.1 Flash for coding agents?
On long agentic tasks, yes, clearly. DeepSeek V4.1 Flash costs under an eighth as much per request in our measurements, so it is the better choice for high-volume routine steps. Many setups use both: Flash for most steps, GLM 5.3 for the hard ones.
Can I use MiniMax M3, Kimi K3 or GLM 5.3 without a Chinese account or phone number?
Yes. On Standard Compute they are minimax/minimax-m3, moonshotai/kimi-k3 and z-ai/glm-5.3 on one key with an OpenAI- and Anthropic-compatible endpoint, from $19 a month. You don't sign up with MiniMax, Moonshot or Z.ai.
Are Qwen 3.7 Plus, Qwen 3.7 Max and Qwen 3.8 Max open-weight?
No. Those are Alibaba's hosted API models. Qwen 3.8 27B is the current open-weight Qwen model we serve, under Apache 2.0.
Is Llama still a good choice for coding agents?
Not in October 2026. Llama 4 Maverick and gpt-oss-120b are easy to run yourself, but for agent work they are far behind GLM 5.3, GLM 5.3 Flash and DeepSeek V4.1 Flash. Pick them for self-hosting constraints, not for agent quality.
Can I run open-weight models in the EU?
Some. With a Europe Compliance rule on Standard Compute, GLM 5.3 and GLM 5.3 Flash run at an EU/EEA endpoint. DeepSeek, Kimi K3, MiniMax M3 and Qwen 3.8 27B have US endpoints only, and the MiMo V2.6 models have no regional endpoint yet; requests to them fail under a Europe rule before any prompt is sent.
Should I self-host an open-weight model for my coding agent?
Only if you have a hard requirement, such as code that may not leave your network. The strong open models need multi-GPU servers, and a hosted API is usually cheaper and faster per agent task. Qwen 3.8 27B is the realistic single-machine option.