Current status: October 1, 2026
For terminal coding agents, start an evaluation with DeepSeek V4.1 Flash. Add GLM-5.3 for text-only engineering tasks, or Kimi K3 when the agent must inspect screenshots or video. For bulk transformations and test drafts, try Flash with non-thinking mode and off-peak scheduling. For self-hosting, compare MIT-licensed DeepSeek weights with Apache-licensed Qwen3-Coder; smaller Gemma 4 variants are a separate option for local code assistance.
Choose by development task
| Coding job | First candidates | Developer constraint |
|---|---|---|
| Terminal agent fixing a repository | V4.1 Flash; GLM-5.3 | Passing tests and tool integration; GLM-5.3 cannot disable reasoning |
| Agent reading UI screenshots | V4.1 Flash; Kimi K3 | Native visual input; test the browser loop |
| Scheduled bulk code changes | V4.1 Flash | Non-thinking mode and off-peak billing |
| Self-hosted code agent | Qwen3-Coder; DeepSeek weights | Serving memory, tool format and weight license |
| Local code drafts | Gemma 4 E2B or E4B | Validate code quality and memory use |
These are evaluation choices based on documented features and billing. This page has no measured cross-provider coding winner or local latency result.
Developer API costs, checked October 1, 2026
| Provider model | Uncached input / output per 1M tokens | Access |
|---|---|---|
| deepseek-flash | $0.30 / $1.20 peak; $0.15 / $0.60 off-peak | DeepSeek API; V4.1 Flash |
| glm-5.3 | $1.40 / $4.40 USD | Z.ai API; Coding Plan has separate quotas |
| kimi-k3 | CNY 20 / CNY 100, published at release | Kimi API; verify the live rate before budgeting |
DeepSeek peak hours are Monday–Friday, 01:00–04:00 and 06:00–10:00 UTC, excluding Chinese public holidays; other hours use the off-peak rate. Cached Flash input costs $0.006 at peak and $0.003 off-peak per million tokens. Z.ai lists $0.26 for cached GLM-5.3 input. Moonshot’s K3 release lists CNY 2 for cached input; its current caching guide lists cache writes separately. Keep currencies and cache charges separate in your estimate.
Use deepseek-flash for new calls: the old V4 Flash identifiers temporarily route to V4.1 Flash. See the DeepSeek migration guide for API details. For other applications, read the general open-weight shortlist.
Official developer sources
Coding specifications and deployment checks, reviewed October 1, 2026
Evaluate a completed code change: can the agent find the relevant files, make the patch and pass the tests? Use Flash for an initial cost baseline. Compare GLM-5.3 or Kimi K3 on difficult tasks, and test downloaded weights when you need to operate inference yourself. Include failed attempts and review time in the cost.
- GLM-5.3 requires reasoning; requests that try to disable it fail.
- Kimi K3 weights are released with multimodal input and a custom license.
- MiniMax M3 weights are available; API rates change above 512K input tokens.
- Kimi K2.7-Code remains a downloadable coding specialist with 256K context.
- The exact agent setup and benchmark version affect coding results.
Compare Completed Repository Tasks
Use the same issue, starting commit, tools and time limit for each candidate. Record whether the patch passes tests, whether it changes unrelated files and how much review it needs. Tool failures and the agent environment affect task completion.
Provider results
Keep benchmark names and versions attached to scores. Different agent setups can produce different results for the same model.
API Limits and Self-Hosting Terms
Published developer specifications
| Model | Context | Weight license | Integration note |
|---|---|---|---|
| V4.1 Flash | 1M API context | MIT | Thinking or non-thinking; tool calls; images |
| V4 Pro | 1M API context | MIT | Continued API service; text input |
| GLM-5.3 | 1M | GLM-5.3 License | Text only; reasoning always enabled |
| Kimi K3 | 1,048,576 | Kimi K3 License | Images and video; terminal-agent use |
| MiniMax M3 | 1M | MiniMax Community License | Multimodal; reasoning can be disabled |
| Kimi K2.7-Code | 256K | Modified MIT | Coding specialist; commercial branding condition |
| Qwen3-Coder-480B-A35B-Instruct | 262,144 native | Apache 2.0 | Non-thinking; model-specific tool format |
Z.ai currently limits API access to the OpenAI Chat Completions protocol for users who have previously held a Coding Plan, including an expired subscription. Check your account’s protocol access before migrating an agent to GLM-5.3.
All listed checkpoints have published weights. Qwen’s named checkpoint has 480B total parameters and Kimi K3 has 2.8T. Active parameters alone do not determine serving memory. Check the weights, context cache and deployment implementation.
License conditions for your coding service
Flash’s MIT license and Qwen’s Apache 2.0 license permit commercial use subject to their terms. Gemma 4 also uses Apache 2.0.
GLM-5.3 requires a security review for model-as-a-service businesses above $10 billion aggregate revenue across the licensee and affiliates over any consecutive 12 months. Kimi K3 requires a separate agreement for model-as-a-service businesses above $20 million aggregate revenue across the licensee and affiliates over any consecutive 12 months. K3 also has branding conditions for large commercial products and exemptions for internal use and official-product or certified-partner access.
M3’s commercial terms require attribution and a one-time notice, or prior written authorization above $20 million yearly revenue for covered products and services. K2.7-Code requires prominent model branding for commercial products above 100 million monthly active users or $20 million monthly revenue.
DeepSeek: Terminal Agents and Bulk Code Work
Try non-thinking Flash for bounded transformations, such as updating repeated API calls or drafting tests. For multi-step terminal tasks, test thinking mode and validate the patch after tool execution. Pro is a fallback candidate when your own evaluation justifies its higher cost.
DeepSeek operating limits
| Item | Current API documentation |
|---|---|
| Flash / Pro identifiers | deepseek-flash / deepseek-v4-pro |
| Maximum output | 384K for both current API models |
| API formats | Chat Completions, Responses and Anthropic |
| FIM completion | Beta; non-thinking mode only |
Source: https://api-docs.deepseek.com/quick_start/pricing/
DeepSeek’s September 10 release reports V4.1 Flash at 90.6 on Terminal-Bench 2.1. Its model card identifies the DeepSeek Harness Minimal setup. Validate results in your chosen agent before using that score to select a model.
Track accepted patches
Include retries, reasoning tokens and human review in the cost. A cheaper call can still produce a more expensive accepted change.
MiniMax M3: Visual Coding and Longer Inputs
MiniMax’s official model card confirms available weights, native image and video input, and local serving instructions. Evaluate it when the agent needs visual feedback during a change.
MiniMax M3 standard API rates
| Input length | Uncached input / output | Cache read |
|---|---|---|
| Up to 512K tokens | $0.30 / $1.20 | $0.06 |
| Above 512K, up to 1M | $0.60 / $2.40 | $0.12 |
Source: https://platform.minimax.io/docs/pricing/overview
Rates are USD per million tokens from MiniMax’s pay-as-you-go page, checked October 1. The release announcement reports 59.0% on SWE-Bench Pro using Claude Code on MiniMax’s infrastructure.
Production access
Compare subscription quotas separately from pay-as-you-go rates. Check that the access plan covers the production requests you intend to send.
Kimi K2.7-Code and K3
K2.7-Code is a coding specialist built on K2.6, with 1T total parameters, 32B active and 256K context. Its downloadable weights include a MoonViT vision encoder. Moonshot reports 62.0 on its own Kimi Code Bench v2.
Kimi coding deployment choices
| Checkpoint | Reason to evaluate | Constraint |
|---|---|---|
| K2.7-Code | An existing coding-specific weight deployment | Modified MIT with branding condition |
| K3 | Longer coding sessions with multimodal context | Larger weights and different license |
Moonshot documents terminal tools and long-horizon coding for K3. Evaluate the benefit on your repository before migrating an existing K2.7-Code deployment. The current sources checked here do not establish a K2.7-Code API token rate.
Inspect provider-published results
Use the benchmark directory to inspect model results and their sources. Check the benchmark version before comparing scores.
Open the benchmark directoryEvaluate finalists on your repository.
Select a Model for Your Coding Job
Developer evaluation steps
- 1Run a terminal issue with Flash, then compare GLM-5.3 or K3 on the same starting commit.
- 2For bulk work, measure non-thinking or lower-effort settings where supported.
- 3Include retries, cache writes and review time in the cost estimate.
- 4For self-hosting, validate memory use, tool format and weight permissions.
- 5Select a candidate after it completes your tasks within the test and review limits.
Acceptance criteria by job
| Job | Measure |
|---|---|
| Terminal fix | Correct patch and passing tests |
| Bulk rewrite | Valid output at the required throughput and total cost |
| Visual agent | Correct behavior after image-based feedback |
| Self-hosted service | Serving latency, memory fit and license terms |
Deployment estimate
A provider’s API price does not estimate your self-hosting cost. Include accelerator usage and operations when comparing the two routes.
For product subscriptions, see the coding tools pricing guide. Broader model choices belong in the task-by-task guide.
Limits of This Comparison
Evaluation boundary
This page checks provider specifications and rates. It has no independently reproduced benchmark, speed test or hardware fit result. Measure task completion before selecting a production model.
Official Coding and Deployment Sources
- DeepSeek Flash benchmark setup and weights and Pro license
- GLM-5.3 developer documentation and license
- Kimi K3 specification and K3 license
- K2.7-Code benchmark and deployment card
- Qwen3-Coder tool format and context
- MiniMax M3 deployment and pricing and plan limits
FAQ
Which open-weight model should I test for terminal coding?
Evaluate DeepSeek V4.1 Flash first, then GLM-5.3 for text-only engineering tasks or Kimi K3 when the agent needs images and video. Use the same repository tasks and agent setup to compare completed changes.
Which model should I use for bulk coding work?
DeepSeek V4.1 Flash is a low-cost starting point for scheduled code transformations and test drafts. Its uncached input/output rates are $0.30/$1.20 at peak and $0.15/$0.60 off-peak per million tokens. Measure retries and review time before choosing it.
Can I self-host these coding models commercially?
Check the exact weight license. DeepSeek V4.1 Flash and V4 Pro use MIT; Qwen3-Coder-480B-A35B-Instruct uses Apache 2.0. GLM-5.3, Kimi K3, MiniMax M3 and Kimi K2.7-Code each have model-specific commercial conditions.
Is MiniMax M3 better than DeepSeek for coding?
This guide has no controlled head-to-head result. MiniMax publishes 59.0% on SWE-Bench Pro, while DeepSeek publishes 90.6 on Terminal-Bench 2.1 for V4.1 Flash. Those different tests cannot establish which model will complete your repository task better.
What context limits matter for coding agents?
DeepSeek’s current API lists 1M context and 384K maximum output. GLM-5.3 lists 1M context and 128K output. Kimi K3 lists 1,048,576 context, while Kimi K2.7-Code and the named Qwen3-Coder checkpoint have 256K native context. Context capacity alone does not prove repository task accuracy.
Get developer model updates
Join the newsletter for model access and pricing updates with links to provider documentation.
Founder of Spectrum AI Labs — testing AI tools and models, and writing up what actually ships.
More about Paras →Stay ahead of the AI curve
We test new AI tools every week and share honest results. Join our newsletter.
