LocalClaw/LLM/Recursive Flint 40B-A2.5B
Open-weight MoE

Recursive Flint 40B-A2.5B

Use case: Coding assistant. Check the memory fit and choose a setup path below.

48 GB short-context planning estimate48 GB RAMQ4_K_MCoding assistant
Choose an app

This model needs a special runtime. Unsupported apps are clearly marked.

Compare models
LM StudioNot available for this model
UnslothNot available for this model
OllamaNot available for this model
Open on Hugging FaceFiles, licence and available downloads
llama.cppNot available for this model
Use with LocalClawOptional workspace after the model is installed
This model needs its official runtimellama.cpp with Flint support (PR 30928, merged as a37cc46104eb50900d5ab4a9ba02a0abdac039db)
Parameters
39.8B (2.5B active, MoE)
RAM planning estimate
48 GB
Model size
23.39 GB
Quantization
Q4_K_M

About Recursive Flint 40B-A2.5B

Catalogue summary: Apache 2.0 coding-agent model trained from scratch by Recursive. ggml-org Q4_K_M GGUF is 23.39GB. Plan for 48GB RAM at short context; 262K is the model limit, not a memory guarantee. Requires a llama.cpp build with Flint support (PR 30928).

Recursive Flint 40B-A2.5B has a 48 GB short-context RAM planning estimate based on its 23.39 GB GGUF. This is not a measured minimum or a guarantee at maximum context.

Use the official llama.cpp with Flint support (PR 30928, merged as a37cc46104eb50900d5ab4a9ba02a0abdac039db) setup. The current low-bit files are not a stock LM Studio install.

Model sourcePublic GGUF repository (verified 2026-11-06): ggml-org/Flint-GGUFUpstream releaseLicence text
codereasoningtool-callinglong-context

Install path

01
Check RAM fitPlan for 48 GB RAM at short context; this is an estimate. Use Q4_K_M.
02
Load the modelFollow the official llama.cpp with Flint support (PR 30928, merged as a37cc46104eb50900d5ab4a9ba02a0abdac039db) instructions. Stock LM Studio support is not confirmed.
03
Try your workloadStart with a short prompt, check memory use and adjust context length before running longer tasks.

Download formats for this model

These are quantizations of the same model. The RAM fit above uses Q4_K_M; larger files need more memory. File sizes exclude the runtime, context cache and optional vision projector.

Catalogue record

  • Family: flint
  • Parameters: 39.8B (2.5B active, MoE)
  • Recommended quantization: Q4_K_M
  • Short-context RAM estimate: 48 GB
  • Catalogue model size: 23.39 GB
  • Tags: code, reasoning, tool-calling, long-context

Practical limits

  • Catalogue RAM is a minimum estimate, not a guarantee for every context length or runtime.
  • Speed and memory use vary by quantization, backend, context length and system headroom.
  • Verify architecture, licence and usage restrictions in the linked upstream material before deployment.

Reading the scores

LocalClaw /10 is an editorial catalogue rating. Community /5 comes from member reviews. Your machine fit is a separate memory and runtime assessment.

Capability profile

Repository catalogue ratings used by LocalClaw's editorial rubric. They are not a standardized third-party benchmark.

speed
?
quality
?
coding
?
reasoning
?

Technical notes

Developer
Recursive
License
Apache 2.0
Context window
262,144 tokens
Architecture
39.8B MoE decoder with approximately 2.5B active parameters, 49 layers, hybrid sliding-window/full attention and learned attention sinks.

This model fits these next steps

Hardware fit is based on LocalClaw's RAM tier, model size and quantization metadata. Always leave memory headroom for your OS and runtime.

Workstation fitMac Studio M4 Max 64GBCUDA workstation classNVIDIA GB10 / DGX SparkLarge local models64GB RAM guide

Related catalogue entries

Linked mechanically by family, shared tags and nearby RAM tier; this is not a quality ranking.

KAT-Coder V2.5 Dev 35B (3B active, MoE)Nemotron 3.5 Lightning 30B-A3B 30B (3B active, MoE)Nanbeige4.3 33B 33B (dense)Nex-N2.5-mini 35B MoEOccamy-1.0 35B (3B active, MoE)Qwen3-Coder-Next 80B (3B active, MoE)Xing4.0-29B-A4B 29B (4B active, MoE)Spark-X2.5-4B 4B

Where to go next