LangGraph
Production orchestration. 256K context plus LangGraph checkpoints handle very long workflows.
Mistral Large 3 is the strongest open-weight model from a Western lab in 2026. Released December 2, 2025 under Apache 2.0, it ships as a 675 billion parameter granular MoE with 41 billion active per token, plus a 2.5 billion parameter vision encoder, plus 256K context.
Specs that matter when picking the model and the hardware.
675B total, 41B active per token, plus 2.5B vision encoder
Granular MoE, multimodal (text + vision)
256K tokens
Apache 2.0
December 2, 2025
3,000 NVIDIA H200 GPUs
73.1 (independent eval)
93.6
~38 tok/s
~340GB, 8x A100/H100/B200
~675GB, 8x H200 node
244GB
Three tiers, real SKUs, real throughput numbers.
Throughput: 10-25 tokens/sec. Best for: Lab and pilot work only, not production serving. Better to use Mistral 3 14B or 8B siblings on a single GPU for SMB.
Throughput: 60-120 tokens/sec. Best for: 20-50 user firms running vision plus text up to 64K context, vLLM with --tensor-parallel-size 4 and the Mistral tool parser. On A100/H100 NVFP4 runs via the Marlin FP4 fallback (memory gain, no speedup over FP8).
Throughput: 150-300 tokens/sec. Best for: Full 256K context production, multi-tenant deployment, agentic workflows with vision.
GGUF, AWQ, FP8, INT4.
NVFP4 for sub-64K-context production on H100 or A100 (near-lossless on text, half the memory of FP8, tensor-parallel-size 4). FP8 for the full 256K context on an 8x H200 node, and FP8 when you plan to fine-tune. Q4 GGUF only for lab and offline batch work.
NVFP4 is calibrated mainly on text, so expect a minor regression on vision datasets and a quality drop past 64K context, which is exactly where you switch to FP8. On A100 or H100 vLLM falls back to Marlin FP4: you keep the memory gain but not the speedup over FP8, and only B200 has the native NVFP4 acceleration path. FP8 is also the more precise choice for fine-tuning. Q4 GGUF works but throughput is bandwidth-bound on consumer GPUs and the tool-calling parser needs Mistral-specific config. The 2.5B vision encoder runs at FP16 regardless of LLM quantization.
File size: ~340GB.
File size: ~675GB.
File size: 244GB.
File size: 370GB.
Released Dec 2 2025 under Apache 2.0, 41B active / 675B total MoE, 256K context, native function calling and JSON output.
Mistral's official agent CLI: npm install -g @mistralai/vibe-cli. Native to the Mistral tool format. Available as a Zed extension. Use this for the smoothest first-party experience, especially when the buyer is already on Mistral's enterprise tier.
Mistral is a first-class Cline provider. OpenAI Compatible config works against your vLLM endpoint. Cline endorsed Devstral 2's tool-calling reliability, and Mistral Large 3 inherits the same chat template. Plan and Act mode functional.
Mistral Large 3 works natively: aider --model mistralai/mistral-large-3 --api-key $MISTRAL_API_KEY for hosted, or --openai-api-base http://your-vllm:8000/v1 for self-hosted. Edit-format applies cleanly across Python, TypeScript, Go.
config.yaml with provider: mistral or openai (custom apiBase). Declare capabilities: [tool_use, image_input] for the multimodal flagship. Roles chat and edit work; pair with a smaller fast model for autocomplete role.
Cline fork. Same OpenAI-compatible config. Multi-mode (Code, Architect, Ask, Debug) functional. Mistral Large 3's 256K context unlocks Architect mode's full repo-context plan.
Works but Devstral 2 is the better Mistral model for repo-autonomy. Use Mistral Large 3 only when the deployment is consolidated to one model. Set LLM_BASE_URL to vLLM, expect SWE-bench Verified pass in the 35 to 45 percent range.
Mistral's enterprise coding bundle: in-IDE assistant, models, local deployment, support. Wraps Mistral Large 3 plus Devstral 2 plus tooling. Used when the buyer wants a vendor-supported package not a stitched stack.
Where this model + agent stack ships real work.
256K context handles deep document synthesis.
Standard OPENAI_BASE_URL override. Mistral Large 3's instruction following is strong, GPT Researcher's plan-and-execute pattern works without prompt tweaking. Reports run 4 to 7 minutes for 5 to 10 sources.
Generic OpenAI provider works. Mistral's parallel_tool_calls flag means @agent mode can call multiple tools per turn, which speeds up multi-tool research workflows.
Document upload, web search, function calling all work. The 256K context is the standout for multi-document research chats. Pair with Mistral's EAGLE speculative decoding for faster streaming.
Production research orchestration. Mistral Large 3's 256K context lets LangGraph hold longer agent state without checkpoint compression. langchain-mistral or langchain-openai with apiBase override both work.
Role-based crews. Mistral Large 3's strong instruction following and parallel tool calls suit CrewAI's sequential pattern. Set llm in agent config.
GroupChat works. The 256K context absorbs more debate history than DeepSeek V3.2's 128K. Mistral's slower output speed (~38 t/s) is the bottleneck on multi-agent debates.
Hybrid stack: LlamaIndex query engines as tools inside a LangGraph or AutoGen workflow. Mistral's strict function-calling mode pairs with LlamaIndex's structured tool definitions cleanly.
Where this model + research agent stack delivers reports your team actually uses.
Pick the framework by workflow shape.
Production orchestration. 256K context plus LangGraph checkpoints handle very long workflows.
Sequential crews with parallel tool calling enabled.
GroupChat. Mistral's slower output speed is the bottleneck on debate-heavy patterns.
Prompt compilation. Strong with Mistral's strict-format following.
Type-safe agents. Mistral Guidance (lark grammar) and Pydantic AI both target structured output.
Retrieval layer. Hybrid stack with LangGraph orchestration.
Mistral's first-party agent CLI with Zed extension.
Three buyer profiles where this model is the right answer.
Task: Multilingual contract review, vision-based document intake (scanned PDFs, exhibits), agentic discovery. Outcome: Apache 2.0 plus Mistral's EU-based provenance simplifies GDPR posture. Vision plus text in one model removes a layer of pipeline glue.
Task: Image-conditioned report drafting, vision-text Q&A on de-identified DICOM exports. Outcome: First-class multimodal in an open-weight model with full HIPAA-compatible self-host. 256K context handles long longitudinal cases.
Task: Document intake (vision OCR plus reasoning), multilingual KYC, agentic workflows with native tool calling. Outcome: Apache 2.0 redistribution lets the firm ship the model embedded in client-facing products without per-seat licensing friction.
Where this model wins, where it loses.
DeepSeek V3.2 wins on hardest reasoning and theorem proving. Mistral Large 3 wins on multimodal (vision is native), Apache 2.0 simplicity (DeepSeek is MIT, comparable but Mistral has stronger Western enterprise familiarity), and 256K context versus 128K.
Qwen3 235B wins on hardware floor (143GB versus 370GB at Q4) and pure reasoning benchmarks. Mistral Large 3 wins on multimodal vision and on training provenance (3,000 H200s, fully disclosed), useful for risk-averse buyers.
Different class. Llama 3.3 70B is single-machine deployable on a $7K Mac. Mistral Large 3 needs an 8x H200 node. Pick Llama for SMB self-host, Mistral Large 3 when frontier quality plus vision plus Apache 2.0 are all required.
Day 4 vLLM serves Mistral Large 3 with the mistral tool-call parser and the three mandatory mistral flags.
vLLM serves Mistral Large 3 with --tokenizer_mode mistral, --config_format mistral, --load_format mistral, --enable-auto-tool-choice, --tool-call-parser mistral, --max-model-len 32768 (raise after memory headroom validated), --tensor-parallel-size 8. mistral_common >= 1.8.6 mandatory or tool calling breaks. NVFP4 for inference, FP8 for fine-tuning.
EAGLE speculative decoding wired with the companion mistralai/Mistral-Large-3-675B-Instruct-2512-Eagle model and num_speculative_tokens: 3. Throughput improves without changing output quality.
Editor agents wired. Aider via --model mistralai/mistral-large-3, Cline as Mistral provider, Continue.dev with provider: mistral, Roo Code OpenAI-compatible. Strict function calling enabled per Mistral 1.10.0 release.
Open WebUI plus AnythingLLM live. Document upload supports image multimodality. Workspaces isolate sensitive doc sets.
LangGraph workflows. Three workflows shipped: deep-research, code-review, customer-intake. Mistral Guidance lark-grammar schema validation wired into the structured-output nodes.
Mistral Code (vendor-supported bundle) or OpenHands repo-autonomy. For coding-heavy buyers we add Devstral 2 alongside Mistral Large 3 because Devstral 2 owns coding (72.2 percent SWE-bench Verified).
Prometheus metrics dashboard (vllm:num_requests_waiting, gpu_cache_usage_perc, time_to_first_token_seconds), evidence pack, runbook, and agent inventory delivered.
Feedback from the people who were in the room when decisions were made.
I used to spend every Sunday reconciling contractor invoices. Now they match automatically. I got my weekends back.
We didn't rush. We just stopped wasting time on things that didn't matter. The risk was managed, not ignored.
The Firebase-to-Supabase migration sounded terrifying. They moved 200k users without a single support ticket. We were live in 19 days.
Fixed fee.
Hardware audit. Confirm 8x H100/H200/B200 node. Pick NVFP4 versus FP8 by context-length need. Run real prompts including vision tasks if applicable.
Serve via vLLM with Mistral tokenizer/config/load formats. Enable tool-call parser ('mistral'). System prompt with tool descriptions. Temperature 0.1 for production.
RAG with vision-capable retrieval if needed. SSO, audit, RBAC. Tune max-model-len to 32K or 64K if you do not need full 256K (saves significant KV-cache memory).
Pen-test cleared. Apache 2.0 NOTICE filed. Production cutover. EU residency confirmed if relevant for GDPR posture.
It means clean redistribution. No MAU cap, no attribution-in-product, no commercial restriction. You preserve the NOTICE file in source distributions, that is the only legal step. You can embed Mistral Large 3 in commercial products, fine-tune it, ship it to clients, with no contact to Mistral required. Counsel typically clears Apache 2.0 in a single review pass.
8x A100 80GB or 8x H100 with NVFP4 weights is the practical floor, $150K to $200K of capex. Below that, you are running Q4 GGUF on multi-GPU consumer cards (4x RTX 5090 or similar) with significant performance compromise. For a firm under 30 users that does not need the full 675B, the smaller siblings (Mistral 3 14B, 8B, 3B) on a single GPU are usually the right answer.
Yes. The 2.5 billion parameter vision encoder is integrated, not bolted on. It handles scanned PDFs, charts, diagrams, photos, and longitudinal multi-image cases. For best results keep aspect ratios near 1:1 and run vision encoder at FP16 regardless of LLM quantization. It is among the first open frontier models with vision and multilingual all in one.
Both are 670B-class MoE on enterprise hardware. Pick Mistral Large 3 if you need vision, longer context (256K versus 128K), or Apache 2.0 plus Western training provenance. Pick DeepSeek V3.2 if you need maximum reasoning and theorem-proving performance and you can live without vision.
Devstral 2 for coding-only deployments. Mistral Large 3 for one-model-many-jobs deployments. Devstral 2 posts 72.2 percent SWE-bench Verified, runs on 4 H100s, and Cline officially endorsed its tool-calling reliability. Mistral Large 3 is the generalist (675B total / 41B active) for teams that want coding plus research plus document Q&A under one license. Pineido often runs both: Mistral Large 3 as the default endpoint, Devstral 2 added on the coding lane when volume justifies it.
Three Mistral-specific flags are mandatory: --tokenizer_mode mistral, --config_format mistral, --load_format mistral. Plus --enable-auto-tool-choice and --tool-call-parser mistral for function calling. Without --tool-call-parser mistral, vLLM uses the default parser and tool calls return empty. Verify mistral_common >= 1.8.6 is installed or tokenization falls back to HuggingFace default and breaks. Pin vLLM >= 0.17.0.
675B total, 41B active per token (MoE), 2.5B vision encoder Parameters
Fixed fee.
Day 1 you see the exact model, the exact hardware, the exact agent stack, the exact cost math, and the exact compliance posture.
ByKostadin StambolievFounder and Imaginator
Pineido, based in Strumica, North Macedonia, is a software company that builds websites, online shops, web and mobile apps, business automation and AI for companies at home and abroad. Work ships in 30-day sprints at a fixed price, and the client owns the code.
Published Updated