Open Model WeightsRECURSIVE · FIELD VERIFIED
Primary sources last checked on 24 Oct 2026. Verification freshness is tracked separately from factual change, and unknown values remain explicit rather than inferred.
FRESHNESS & VERIFICATION
A verification check does not pretend the model changed. Open Model Weights updates the verification date after a real primary-source check, while factual-update dates move only when retained source evidence changes.
CONFLICT INTELLIGENCE
+ Save to my shortlistMonitor changes →Local browser monitoring · no background email alerts
DIRECT ANSWER · FEATURED MODEL BRIEFING
Source-grounded editorial layer · reviewed 2026-10-23
Recursive Flint is a downloadable Apache-2.0 coding-agent mixture-of-experts model with approximately 39.8 billion total parameters and 2.5 billion parameters active per token, according to Recursive. The source documents text reasoning, multi-turn tool calls and a 262,144-token configured context, with BF16 weights and an SGLang serving kit. The 2.5B active figure is not the model's stored weight size or a VRAM measurement. Recursive's Terminal-Bench and SWE-bench scores remain publisher-reported. Open Model Weights classifies Flint as Notable pending independent comparative testing and will only claim source-field verification after checking the checkpoint's exact files, current SHA and Apache-2.0 declaration.
WHY IT MATTERS
Sparse activation offers a different compute profile for coding agents but leaves full parameter residency as a separate hardware consideration.
Downloadable weights and an explicit serving guide make it feasible to design a versioned, source-grounded coding agent benchmark.
MODEL-SELECTION VIEW
Recursive Flint is described by its publisher as a decoder-only sparse mixture-of-experts model with 39.8 billion total parameters and approximately 2.5 billion active per token. Its 49 transformer layers include 2 dense layers and 47 MoE layers; each MoE block routes tokens through six of 256 experts plus a shared expert. The lower active count describes one step of computation, not the full checkpoint that must be stored or accessed.
Flint combines 35 sliding-window attention layers, with a 513-token local window, and 14 full-attention layers. The model card specifies 32 query heads, four key/value heads and a 262,144-token context configuration shared by inputs and generated outputs. These are published architectural characteristics, not a guarantee of robust reasoning over every length or an estimate of usable KV-memory on a particular device.
The source describes Flint as a model for inspecting repositories, finding defects, implementing features and writing tests through multistep agentic workflows. It supports text-in/text-out reasoning and tool calls, so successful use requires a compatible chat template and orchestration loop that can receive tool results and preserve prior reasoning state in the expected message format.
An agent benchmark is not just a model benchmark. Shell tools, patch application, retries, allowed network access and the quality of the error messages all affect whether a candidate patch passes tests. Teams should isolate Flint from secrets, prohibit privileged tool calls by default and evaluate failure cases, including repository instructions that deliberately try to redirect the agent.
Recursive reports 54.9 percent on Terminal-Bench 2.1, 89.7 percent on SWE-bench Pro v2 and 74.2 percent on SWE-bench Verified. The model card specifies sampling at temperature 1.0, top-p 0.95 and top-k 20; agentic scores are aggregated across at least five attempts. Results marked by the publisher are not Open Model Weights independent benchmark executions.
The Terminal-Bench release methodology counts answers obtained through web retrieval as unsuccessful, but other teams may use different rules. Mixing numbers across test harness versions or different web/network privileges can imply a nonexistent head-to-head advantage. Before declaring a model Leading, our registry would require externally comparable tests, stable task definitions and evidence that the exact public checkpoint was used.
The official Flint repository contains the BF16 checkpoint, tokenizer, configuration, chat template and an SGLang serving kit. The publisher shows single-replica serving on an H200 or B300, while separate quantized ecosystem variants target other hardware, including Apple Silicon. The original BF16 repository is not interchangeable with a third-party GGUF or low-bit checkpoint.
The downloadable BF16 archive is approximately 80 GB according to the publisher's getting-started example. That does not include all inference allocations, token cache, activations, concurrent sessions or model-server processes. Profile memory, time to first token and task-completion latency with real repository sizes; token-per-second figures for short prompts cannot alone establish feasible production serving.
The Hugging Face model card declares Apache 2.0 for Flint's released weights. Apache's redistribution and attribution obligations should be reviewed alongside the specific serving container dependencies, plugins and third-party quantized builds. Before production, record the repository SHA and the verified weight files rather than linking only to a moving main branch.
Generated code can contain logic defects, security vulnerabilities or malicious changes even if the model is permissively licensed. A responsible deployment keeps CI secrets outside the agent environment, requires human approval for destructive commands and merges, and evaluates the model on malicious tool outputs as well as ordinary coding tasks.
Flint is Notable for a genuinely downloadable, specialized coding-agent design with a highly sparse parameter activation profile and detailed first-party serving instructions. Its claims are relevant to smaller GPU deployments and self-hosted development assistants, but the public release alone does not establish a top-tier rank across independent intelligence tests.
The practical decision is whether Flint solves your own engineering issues more accurately per total task minute, not whether its active-parameter number is smaller than a competing dense model's size. Use a fixed private task suite, collect patch acceptance and regression failures, measure the whole serving stack and keep a tested fallback model until the improvement holds across representative projects.
IMPLEMENTATION & EVALUATION
Four-stage checkpoint-pinned coding-agent evaluation with safe tools and independent task metrics; no claim of execution by OMW.
Capture the exact official repository SHA, required tokenizer/chat template, model files, license notice and tested SGLang images.
A reviewed immutable manifest, smoke test and single working tool-call conversation.
Changing a template or quantized build invalidates an apparent head-to-head result.
Put execution behind isolated containers, a write allowlist, network restrictions and clean repository fixtures.
No secret reads or unreviewed privileged operations are possible; all tool calls logged.
Repository prompt injection can redirect otherwise correct tool-use policies.
Benchmark bug fixes, code search, refactors and unit tests with the same harness and budgets used for a baseline.
For each task, publish verified tests passed, human interventions, retries and elapsed time.
Private cherry-picked examples or retries inflate pass rate.
Profile realistic context lengths and concurrent sessions at chosen BF16 or quantized runtime settings.
Peak memory, end-to-end task throughput, failure rate and rollback thresholds documented.
Short synthetic decode rates conceal orchestration and context cost.
Proposed tests, not OMW benchmark results. Read the official model card at the pinned source revision ↗ Inspect OMW field evidence ↗
PRIMARY & INDEPENDENT SOURCES
FAQ
Recursive publishes a downloadable BF16 checkpoint under an Apache-2.0 declaration, together with model configuration and a serving kit. Exact revision and file availability require canonical verification.
The first figure counts overall model parameters while the second describes the approximate parameters active for one token. Installed weights and GPU memory cannot be calculated from 2.5B alone.
The publisher provides SGLang recipes for one H200 or B300 per replica. Hardware compatibility is deployment-specific, and Open Model Weights has not independently executed those recipes.
The model card describes a 262,144-token configuration shared between prompt and output. Quality at extreme lengths and actual serving memory are not established by that configuration value.
Not by this site. Recursive provides benchmark results with stated sampling and evaluation conditions; a valid comparison needs the same harness and the exact released checkpoint.
Use a sandboxed, permission-limited task runner on a frozen repository benchmark with test logs, review of generated diffs, measured full-task latency and the official checkpoint SHA.
MODEL INTELLIGENCE · SOURCE-FIRST
Open Model Weights records Recursive Flint 40B-A2.5B as a 39.7853B-parameter text/code to text; tool-using coding agent open-weight model in the Flint family, with Recursive listed as the developer. The checked source repository provides 16 recognized weight artifacts, format evidence for Safetensors and a verified context window of 262,144 tokens. Values that are not evidenced in the checked sources remain explicitly undisclosed or unknown.
The repository declares apache-2.0. Commercial use allowed under apache-2.0 conditions. Open Model Weights reports the checked license evidence and classification; this is not legal advice.
No base model is declared in the checked standard lineage metadata, so Open Model Weights does not infer one. Training recipe status: training-described-in-model-card; data disclosure status: not-disclosed-in-standard-metadata-or-obvious-model-card-section.
DEPLOYMENT PROFILE
These bars visualize weight-only decimal-GB estimates. They exclude KV cache, activations, runtime overhead, optimizer state and sharding.
LICENSE & USE
Commercial use allowed under apache-2.0 conditions
This section summarizes the checked repository evidence. It is not legal advice. Inspect license evidence ↗
LINEAGE & PROVENANCE
Open Model Weights does not infer a hidden base model or training corpus when the checked sources do not declare one.
RUNTIME SIGNALS
Declared and model-card-mentioned runtimes are discovery signals. They are not treated as Open Model Weights runtime tests unless explicitly measured.
EVIDENCE TRAIL
This is where the article remains auditable: source links on one side, the registry's observed change history on the other.
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Checked Hugging Face API metadata, exact repository file list, model card and config when accessible.
01 · WEIGHTS
model-00001-of-00015.safetensorsmodel-00002-of-00015.safetensorsmodel-00003-of-00015.safetensorsmodel-00004-of-00015.safetensorsmodel-00005-of-00015.safetensorsmodel-00006-of-00015.safetensorsmodel-00007-of-00015.safetensorsmodel-00008-of-00015.safetensorsmodel-00009-of-00015.safetensorsmodel-00010-of-00015.safetensorsmodel-00011-of-00015.safetensorsmodel-00012-of-00015.safetensorsmodel-00013-of-00015.safetensorsmodel-00014-of-00015.safetensorsmodel-00015-of-00015.safetensorsmodel.safetensors.index.json02 · LICENSE
Classification describes the checked license terms; it is not legal advice.
License / model-card evidence ↗03 · HARDWARE
Weight-only decimal-GB estimate from total parameter count; excludes KV cache, activations, runtime overhead, optimizer state and sharding.
04 · FORMATS
“Not observed” means no matching repository artifact/name/dtype signal was found; it does not prove no third-party conversion exists.
05 · LINEAGE
06 · TRAINING ASSETS
07 · RUNTIME SUPPORT
Open Model Weights separates model-repository evidence from external runtime-release evidence. An external support statement does not change the official model repository's weight formats and does not imply an Open Model Weights inference benchmark.
08 · SOURCES
09 · VERIFICATION HISTORY
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Official Hugging Face repository SHA unchanged; prior field evidence retained.
Checked Hugging Face API metadata, exact repository file list, model card and config when accessible.
MODELS FOR AI AGENTS · ROLE EVIDENCE
Specialized model. A supporting component. This checkpoint does not, by category alone, serve as the planning and tool-execution controller.
The curated notable tier is not an agent-readiness rating. Model-specific tool-call success, long-horizon task accuracy and compatible runtime configurations are not independently established here.
Source-linked license: apache-2.0 (allowed commercial-use classification; see original terms).
Context: 262144 configured tokens.
BF16 weights: 79.57 estimated decimal GB, not runtime VRAM.
Compare agent model candidates → · Agent evaluation method → · Inspect original model repository ↗
These records are separately sourced. They do not replace independent model-plus-runtime tests.
No model-specific agent capability evaluation is recorded for this checkpoint. Missing evidence is unknown, not a failed task.
Read evidence details and benchmarks → · Use the agent model selector →