Tasks and inputs
Sarvam-105B-1711 is listed for Text generation, Text → Text. It accepts text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
Sarvam-105B-1711 is a refreshed open post-train of Sarvam's 105B MoE, tuned for tool calling, instruction following, agents and Indian languages.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Pricing conditions: No checkpoint-specific price. The hosted sarvam-105b API is INR 10 in / INR 0.25 cached / INR 25 out per 1M tokens (introductory, through end of 2026), but it is not confirmed to serve 1711. Indicative.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
Same base, new post-training. Sarvam-105B-1711 takes the existing Sarvam 105B mixture-of-experts (105B total, 10.3B active) and retrains the later stages for four things: tool calling, instruction following, agentic tasks and Indian languages. The repository appeared on 15 November 2026, without a public announcement alongside it.
So what is actually new? Mostly the scores and the packaging.
| Benchmark | BF16 | NVFP4 |
|---|---|---|
| IFEval | 96.5 | 96.4 |
| tau2-Bench | 85.5 | 84.1 |
| AIME 2025 | 87.9 | 86.0 |
| GPQA Diamond | 80.4 | 78.2 |
| BrowseComp | 41.4 | 35.0 |
| Indic Suite | 79.1 | 78.8 |
IFBench is reported at 82.5. Most quantised scores sit within a point or two of BF16. BrowseComp is the exception, a six-point drop, so if web-browsing agents are your use case, evaluate the BF16 weights before settling on the smaller build. All figures are publisher-reported.
Three precisions exist. The main repository is NVFP4; BF16 weights live in sarvamai/sarvam-105b-1711-bf16; an MXFP8 build is tagged too. The published vLLM recipe serves NVFP4 on a single B200 at 131,072 max length, reporting about 532 tokens/s for one user and about 7,490 tokens/s total at 128 users. A single 96GB RTX PRO 6000 also works, at a 50K context and roughly 31 tokens/s.
The recipe uses an FP8 KV cache and GLM-4.5-style tool and reasoning parsers. It also ships a DSpark speculative-decoding head, which drafts five tokens ahead and requires the separate draft model in the download directory. The throughput figures above include it.
enable_thinking: False in the chat-template arguments.Sarvam's hosted API lists sarvam-105b and sarvam-105b-conversations, with no dated 1711 version. Whether those endpoints already run this checkpoint is not stated. The API's introductory rate for sarvam-105b is INR 10 in, INR 0.25 cached and INR 25 out per 1M tokens through the end of 2026, but do not assume you are getting 1711 at that price. If the distinction matters, self-host or ask Sarvam support which build sits behind the endpoint.
One more oddity: the model page shows a 56B parameter figure next to the 105B total. That is a display artefact of the packed 4-bit format rather than a different model. The "1711" suffix reads like a 17 November date stamp even though the repository is dated two days earlier.
If you already run the original 105B, swap checkpoints in a staging environment and replay a week of real tool-calling traffic. Compare call validity and task completion, not just benchmark deltas. Teams starting fresh with Indic workloads should also try the smaller Sarvam 30B for cost.
Compare the model on the work you need to complete.
Sarvam-105B-1711 is listed for Text generation, Text → Text. It accepts text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.
The weights are downloadable under Apache-2.0. Check runtime support and size the complete model and working context for your expected concurrency.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
Tencent
Tencent Hy4 Flash is the cost-focused companion to Hy4, keeping vision and agent skills at a far lower token price. Availability is still unsettled.
Alibaba
Qwen4 Max (qwen4-max) is Alibaba's first Qwen4-generation flagship: 1M-token context, text, image and video input, thinking on by default, $1.50/$4.50 per 1M.
Alibaba
Qwen4 Max Preview was Alibaba's five-day early-access build of its Qwen4 flagship, offered only through a Token Plan subscription before GA qwen4-max.
Gemini 4 Carbon is Google's strongest Gemini 4 reasoning model for long-horizon agentic coding, scientific research and cyber defence, with a 1M context.
Anthropic
Claude Fable 5.2 is Anthropic's top model for demanding reasoning and long-horizon agentic work, with a 1M context at $8/$40, 20% below Fable 5.1.
Anthropic
Claude Mythos 5.2 is the verified-access edition of Claude Fable 5.2 for vetted cyber and life-sciences organisations, with a 1M context at $8/$40.