Available Models

FreeInference exposes an OpenAI-compatible model catalog for coding agents and IDEs. The catalog changes as providers and local deployments change, so the authenticated API is the source of truth for your account:

curl -H "Authorization: Bearer $FREEINFERENCE_API_KEY" \
  https://freeinference.org/v1/models

The response is filtered by access level. A model that is not returned for your key is not available to that account.

Model Overview

Model ID

Access

Context

Max Output

Input

Highlights

glm-5.3-flash

Free

1M

131K

Text, image

Coding and agentic work, tools, structured output, always-on thinking with reasoning_effort, tool streaming

deepseek-v4-flash

Free

1M

393K

Text

Agentic coding, tools, structured output, reasoning controls

minimax-m3

Free

1M

131K

Text, image, video

Long context, tools, structured output, thinking

qwen3.6-35b

Free

262K

8K

Text, image, video

Dual local NVFP4/FP8 replicas, fast non-thinking model, tools, structured output

glm-5.2

Pro

1M

131K

Text

Tools, structured output, thinking, tool streaming

kimi-k2.7-code

Pro

262K

131K

Text, image, video

Coding agents, tools, structured output, thinking

All listed models produce text. Context and output limits are deployment limits; an upstream provider may enforce a smaller limit for an individual request.

DeepSeek Upstream Reference Pricing

The following prices are DeepSeek’s upstream API reference rates in U.S. dollars per 1 million tokens. The gateway uses them for cost accounting and route selection; they are not fees charged by FreeInference to users. See the official DeepSeek pricing page for the latest rates.

Since 2026-08-16 16:00 UTC, DeepSeek itself bills at peak and off-peak rates: peak hours are 01:00–04:00 UTC and 06:00–10:00 UTC each day, and off-peak is half the peak rate.

FreeInference does not track that split. The gateway accounts for every DeepSeek call at the peak rate, whatever the hour:

Model

Input (cache hit)

Input (cache miss)

Output

V4 Flash (deepseek-v4-flash)

$0.014

$0.44

$1.32

Using the peak rate uniformly means an off-peak call is accounted at twice the rate DeepSeek actually charged for it. These figures drive internal cost accounting and route selection; they are not fees billed to users, so the effect is on quota consumption, not on anything you pay.

Embedding Model

Model ID

Access

Context

Input

Output

bge-m3

Free

8K

Text

Embedding vector

Use bge-m3 with the OpenAI-compatible /v1/embeddings endpoint for codebase indexing. It is not a chat or completion model.

Choosing a Model

  • Start with glm-5.3-flash for general coding and agentic work: 1M context and always-on reasoning with reasoning_effort control.

  • Use qwen3.6-35b for quick edits and background-agent calls that do not need extended reasoning.

  • Use minimax-m3 for long-context or multimodal work.

  • Use deepseek-v4-flash for complex agentic coding and long reasoning chains.

  • Pro users can add glm-5.2 when thinking has to be switchable (same 1M context), or kimi-k2.7-code for coding-agent workflows.

  • glm-5.3-flash always reasons: thinking: {"type": "disabled"} is ignored rather than honored. Send reasoning_effort (low, high, max) to control how much it thinks, and use glm-5.2 if your client needs thinking fully off.

Access and Retired IDs

Free accounts can use models marked Free. Models marked Pro require a Pro-enabled key. Operational models restricted to administrators or staff are intentionally omitted from the public recommendation list.

Old IDs such as glm-4.7 and minimax-m2.7 are no longer in the production catalog. If a saved IDE configuration uses a retired ID, select one returned by GET /v1/models; otherwise the gateway returns 404 Model not found.

A 404 does not always mean retirement: the same error is returned for a model that exists but is not available at your key’s access level. Either way, the models listed by GET /v1/models for your key are the ones it can use.

Switching Models

Use the exact model ID from GET /v1/models in your client’s model selector or configuration. In Claude Code, use /model; for provider-specific setup, see the IDE and coding-agent integrations.